File size: 176,776 Bytes
e845c9c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 | program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
{
func main<ios18>(tensor<int32, [1, 192]> token_ids, tensor<int32, [1]> token_len) {
tensor<int32, [1, 192]> var_37 = const()[name = string("op_37"), val = tensor<int32, [1, 192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
tensor<int32, [1]> seq_length_expand_1_axes_0 = const()[name = string("seq_length_expand_1_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<int32, [1, 1]> seq_length_expand_1 = expand_dims(axes = seq_length_expand_1_axes_0, x = token_len)[name = string("seq_length_expand_1")];
tensor<bool, [1, 192]> var_43 = greater_equal(x = var_37, y = seq_length_expand_1)[name = string("op_43")];
tensor<bool, [1, 192]> var_44 = logical_not(x = var_43)[name = string("op_44")];
tensor<int32, [1]> valid_axes_0 = const()[name = string("valid_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<bool, [1, 192, 1]> valid = expand_dims(axes = valid_axes_0, x = var_44)[name = string("valid")];
int32 var_50_axis_0 = const()[name = string("op_50_axis_0"), val = int32(0)];
int32 var_50_batch_dims_0 = const()[name = string("op_50_batch_dims_0"), val = int32(0)];
bool var_50_validate_indices_0 = const()[name = string("op_50_validate_indices_0"), val = bool(false)];
tensor<fp16, [6561, 512]> input_embedding_weight_to_fp16 = const()[name = string("input_embedding_weight_to_fp16"), val = tensor<fp16, [6561, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(896)))];
string token_ids_to_int16_dtype_0 = const()[name = string("token_ids_to_int16_dtype_0"), val = string("int16")];
tensor<int16, [1, 192]> token_ids_to_int16 = cast(dtype = token_ids_to_int16_dtype_0, x = token_ids)[name = string("cast_106")];
tensor<fp16, [1, 192, 512]> var_50_cast_fp16_cast_uint16 = gather(axis = var_50_axis_0, batch_dims = var_50_batch_dims_0, indices = token_ids_to_int16, validate_indices = var_50_validate_indices_0, x = input_embedding_weight_to_fp16)[name = string("op_50_cast_fp16_cast_uint16")];
string cast_4_to_fp16_dtype_0 = const()[name = string("cast_4_to_fp16_dtype_0"), val = string("fp16")];
tensor<fp16, [1, 192, 1]> valid_to_fp16 = cast(dtype = cast_4_to_fp16_dtype_0, x = valid)[name = string("cast_105")];
tensor<fp16, [1, 192, 512]> xs_1_cast_fp16 = mul(x = var_50_cast_fp16_cast_uint16, y = valid_to_fp16)[name = string("xs_1_cast_fp16")];
fp32 var_69 = const()[name = string("op_69"), val = fp32(0x1.47ae14p-7)];
int32 var_77 = const()[name = string("op_77"), val = int32(-1)];
tensor<int32, [1]> var_118_axes_0 = const()[name = string("op_118_axes_0"), val = tensor<int32, [1]>([1])];
tensor<bool, [1, 1, 192]> var_118 = expand_dims(axes = var_118_axes_0, x = var_43)[name = string("op_118")];
tensor<bool, [1, 1, 192]> masks_1 = logical_not(x = var_118)[name = string("masks_1")];
tensor<fp16, [512, 512]> encoder_embed_out_0_weight_to_fp16 = const()[name = string("encoder_embed_out_0_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6719424)))];
tensor<fp16, [512]> encoder_embed_out_0_bias_to_fp16 = const()[name = string("encoder_embed_out_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7243776)))];
tensor<fp16, [1, 192, 512]> linear_0_cast_fp16 = linear(bias = encoder_embed_out_0_bias_to_fp16, weight = encoder_embed_out_0_weight_to_fp16, x = xs_1_cast_fp16)[name = string("linear_0_cast_fp16")];
tensor<int32, [1]> input_3_axes_0 = const()[name = string("input_3_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [512]> encoder_embed_out_1_weight_to_fp16 = const()[name = string("encoder_embed_out_1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7244864)))];
tensor<fp16, [512]> encoder_embed_out_1_bias_to_fp16 = const()[name = string("encoder_embed_out_1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7245952)))];
fp16 var_72_to_fp16 = const()[name = string("op_72_to_fp16"), val = fp16(0x1.5p-17)];
tensor<fp16, [1, 192, 512]> input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = encoder_embed_out_1_bias_to_fp16, epsilon = var_72_to_fp16, gamma = encoder_embed_out_1_weight_to_fp16, x = linear_0_cast_fp16)[name = string("input_3_cast_fp16")];
fp16 var_131_to_fp16 = const()[name = string("op_131_to_fp16"), val = fp16(0x1.6ap+4)];
tensor<fp16, [1, 192, 512]> x_3_cast_fp16 = mul(x = input_3_cast_fp16, y = var_131_to_fp16)[name = string("x_3_cast_fp16")];
tensor<int32, [1]> var_158_axes_0 = const()[name = string("op_158_axes_0"), val = tensor<int32, [1]>([-1])];
bool var_158_keep_dims_0 = const()[name = string("op_158_keep_dims_0"), val = bool(true)];
string cast_6_to_fp16_dtype_0 = const()[name = string("cast_6_to_fp16_dtype_0"), val = string("fp16")];
tensor<fp16, [1, 1, 192]> masks_1_to_fp16 = cast(dtype = cast_6_to_fp16_dtype_0, x = masks_1)[name = string("cast_104")];
tensor<fp16, [1, 1, 1]> var_158_cast_fp16 = reduce_sum(axes = var_158_axes_0, keep_dims = var_158_keep_dims_0, x = masks_1_to_fp16)[name = string("op_158_cast_fp16")];
fp16 var_79_promoted_to_fp16 = const()[name = string("op_79_promoted_to_fp16"), val = fp16(0x0p+0)];
tensor<bool, [1, 1, 1]> all_masked_1_cast_fp16 = equal(x = var_158_cast_fp16, y = var_79_promoted_to_fp16)[name = string("all_masked_1_cast_fp16")];
fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)];
tensor<fp16, [1, 1, 192]> fill_like_0_cast_fp16 = fill_like(ref_tensor = masks_1, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")];
tensor<int32, [3]> all_masked_1_after_broadcast_reps_0 = const()[name = string("all_masked_1_after_broadcast_reps_0"), val = tensor<int32, [3]>([1, 1, 192])];
tensor<bool, [1, 1, 192]> all_masked_1_after_broadcast = tile(reps = all_masked_1_after_broadcast_reps_0, x = all_masked_1_cast_fp16)[name = string("all_masked_1_after_broadcast")];
tensor<fp16, [1, 1, 192]> mask_1_cast_fp16 = select(a = fill_like_0_cast_fp16, b = masks_1_to_fp16, cond = all_masked_1_after_broadcast)[name = string("mask_1_cast_fp16")];
tensor<int32, [3]> var_164_perm_0 = const()[name = string("op_164_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
tensor<int32, [6]> input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 0, 3])];
string input_9_mode_0 = const()[name = string("input_9_mode_0"), val = string("constant")];
fp16 const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = fp16(0x0p+0)];
tensor<fp16, [1, 512, 192]> var_164_cast_fp16 = transpose(perm = var_164_perm_0, x = x_3_cast_fp16)[name = string("transpose_104")];
tensor<fp16, [1, 512, 195]> input_9_cast_fp16 = pad(constant_val = const_9_to_fp16, mode = input_9_mode_0, pad = input_9_pad_0, x = var_164_cast_fp16)[name = string("input_9_cast_fp16")];
string input_11_pad_type_0 = const()[name = string("input_11_pad_type_0"), val = string("valid")];
tensor<int32, [1]> input_11_strides_0 = const()[name = string("input_11_strides_0"), val = tensor<int32, [1]>([1])];
tensor<int32, [2]> input_11_pad_0 = const()[name = string("input_11_pad_0"), val = tensor<int32, [2]>([0, 0])];
tensor<int32, [1]> input_11_dilations_0 = const()[name = string("input_11_dilations_0"), val = tensor<int32, [1]>([1])];
int32 input_11_groups_0 = const()[name = string("input_11_groups_0"), val = int32(1)];
tensor<fp16, [512, 512, 4]> encoder_pre_lookahead_layer_conv1_weight_to_fp16 = const()[name = string("encoder_pre_lookahead_layer_conv1_weight_to_fp16"), val = tensor<fp16, [512, 512, 4]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7247040)))];
tensor<fp16, [512]> encoder_pre_lookahead_layer_conv1_bias_to_fp16 = const()[name = string("encoder_pre_lookahead_layer_conv1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9344256)))];
tensor<fp16, [1, 512, 192]> input_11_cast_fp16 = conv(bias = encoder_pre_lookahead_layer_conv1_bias_to_fp16, dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = encoder_pre_lookahead_layer_conv1_weight_to_fp16, x = input_9_cast_fp16)[name = string("input_11_cast_fp16")];
tensor<fp16, [1, 512, 192]> input_13_cast_fp16 = leaky_relu(alpha = var_69, x = input_11_cast_fp16)[name = string("input_13_cast_fp16")];
tensor<int32, [6]> input_15_pad_0 = const()[name = string("input_15_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 2, 0])];
string input_15_mode_0 = const()[name = string("input_15_mode_0"), val = string("constant")];
fp16 const_10_to_fp16 = const()[name = string("const_10_to_fp16"), val = fp16(0x0p+0)];
tensor<fp16, [1, 512, 194]> input_15_cast_fp16 = pad(constant_val = const_10_to_fp16, mode = input_15_mode_0, pad = input_15_pad_0, x = input_13_cast_fp16)[name = string("input_15_cast_fp16")];
string outputs_1_pad_type_0 = const()[name = string("outputs_1_pad_type_0"), val = string("valid")];
tensor<int32, [1]> outputs_1_strides_0 = const()[name = string("outputs_1_strides_0"), val = tensor<int32, [1]>([1])];
tensor<int32, [2]> outputs_1_pad_0 = const()[name = string("outputs_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
tensor<int32, [1]> outputs_1_dilations_0 = const()[name = string("outputs_1_dilations_0"), val = tensor<int32, [1]>([1])];
int32 outputs_1_groups_0 = const()[name = string("outputs_1_groups_0"), val = int32(1)];
tensor<fp16, [512, 512, 3]> encoder_pre_lookahead_layer_conv2_weight_to_fp16 = const()[name = string("encoder_pre_lookahead_layer_conv2_weight_to_fp16"), val = tensor<fp16, [512, 512, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9345344)))];
tensor<fp16, [512]> encoder_pre_lookahead_layer_conv2_bias_to_fp16 = const()[name = string("encoder_pre_lookahead_layer_conv2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10918272)))];
tensor<fp16, [1, 512, 192]> outputs_1_cast_fp16 = conv(bias = encoder_pre_lookahead_layer_conv2_bias_to_fp16, dilations = outputs_1_dilations_0, groups = outputs_1_groups_0, pad = outputs_1_pad_0, pad_type = outputs_1_pad_type_0, strides = outputs_1_strides_0, weight = encoder_pre_lookahead_layer_conv2_weight_to_fp16, x = input_15_cast_fp16)[name = string("outputs_1_cast_fp16")];
tensor<int32, [3]> var_185_perm_0 = const()[name = string("op_185_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
tensor<fp16, [1, 192, 512]> var_185_cast_fp16 = transpose(perm = var_185_perm_0, x = outputs_1_cast_fp16)[name = string("transpose_103")];
tensor<fp16, [1, 192, 512]> input_17_cast_fp16 = add(x = var_185_cast_fp16, y = x_3_cast_fp16)[name = string("input_17_cast_fp16")];
tensor<int32, [1]> query_1_axes_0 = const()[name = string("query_1_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [512]> encoder_encoders_0_norm_mha_weight_to_fp16 = const()[name = string("encoder_encoders_0_norm_mha_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10919360)))];
tensor<fp16, [512]> encoder_encoders_0_norm_mha_bias_to_fp16 = const()[name = string("encoder_encoders_0_norm_mha_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10920448)))];
fp16 var_58_to_fp16 = const()[name = string("op_58_to_fp16"), val = fp16(0x1p-24)];
tensor<fp16, [1, 192, 512]> query_1_cast_fp16 = layer_norm(axes = query_1_axes_0, beta = encoder_encoders_0_norm_mha_bias_to_fp16, epsilon = var_58_to_fp16, gamma = encoder_encoders_0_norm_mha_weight_to_fp16, x = input_17_cast_fp16)[name = string("query_1_cast_fp16")];
tensor<fp16, [512, 512]> encoder_encoders_0_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_encoders_0_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10921536)))];
tensor<fp16, [512]> encoder_encoders_0_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_encoders_0_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11445888)))];
tensor<fp16, [1, 192, 512]> linear_1_cast_fp16 = linear(bias = encoder_encoders_0_self_attn_linear_q_bias_to_fp16, weight = encoder_encoders_0_self_attn_linear_q_weight_to_fp16, x = query_1_cast_fp16)[name = string("linear_1_cast_fp16")];
tensor<int32, [4]> var_207 = const()[name = string("op_207"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 192, 8, 64]> q_1_cast_fp16 = reshape(shape = var_207, x = linear_1_cast_fp16)[name = string("q_1_cast_fp16")];
tensor<fp16, [512, 512]> encoder_encoders_0_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_encoders_0_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11446976)))];
tensor<fp16, [512]> encoder_encoders_0_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_encoders_0_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11971328)))];
tensor<fp16, [1, 192, 512]> linear_2_cast_fp16 = linear(bias = encoder_encoders_0_self_attn_linear_k_bias_to_fp16, weight = encoder_encoders_0_self_attn_linear_k_weight_to_fp16, x = query_1_cast_fp16)[name = string("linear_2_cast_fp16")];
tensor<int32, [4]> var_212 = const()[name = string("op_212"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 192, 8, 64]> k_1_cast_fp16 = reshape(shape = var_212, x = linear_2_cast_fp16)[name = string("k_1_cast_fp16")];
tensor<fp16, [512, 512]> encoder_encoders_0_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_encoders_0_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11972416)))];
tensor<fp16, [512]> encoder_encoders_0_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_encoders_0_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12496768)))];
tensor<fp16, [1, 192, 512]> linear_3_cast_fp16 = linear(bias = encoder_encoders_0_self_attn_linear_v_bias_to_fp16, weight = encoder_encoders_0_self_attn_linear_v_weight_to_fp16, x = query_1_cast_fp16)[name = string("linear_3_cast_fp16")];
tensor<int32, [4]> var_217 = const()[name = string("op_217"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 192, 8, 64]> v_1_cast_fp16 = reshape(shape = var_217, x = linear_3_cast_fp16)[name = string("v_1_cast_fp16")];
tensor<int32, [4]> v_3_perm_0 = const()[name = string("v_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<fp16, [8, 64]> const_13_to_fp16 = const()[name = string("const_13_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12497856)))];
tensor<fp16, [1, 192, 8, 64]> var_230_cast_fp16 = add(x = q_1_cast_fp16, y = const_13_to_fp16)[name = string("op_230_cast_fp16")];
tensor<fp16, [8, 64]> const_14_to_fp16 = const()[name = string("const_14_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12498944)))];
tensor<fp16, [1, 192, 8, 64]> var_233_cast_fp16 = add(x = q_1_cast_fp16, y = const_14_to_fp16)[name = string("op_233_cast_fp16")];
tensor<int32, [4]> q_with_bias_v_1_perm_0 = const()[name = string("q_with_bias_v_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
bool matrix_ac_1_transpose_x_0 = const()[name = string("matrix_ac_1_transpose_x_0"), val = bool(false)];
bool matrix_ac_1_transpose_y_0 = const()[name = string("matrix_ac_1_transpose_y_0"), val = bool(false)];
tensor<int32, [4]> transpose_30_perm_0 = const()[name = string("transpose_30_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
tensor<int32, [4]> transpose_31_perm_0 = const()[name = string("transpose_31_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
tensor<fp16, [1, 8, 64, 192]> transpose_31 = transpose(perm = transpose_31_perm_0, x = k_1_cast_fp16)[name = string("transpose_99")];
tensor<fp16, [1, 8, 192, 64]> transpose_30 = transpose(perm = transpose_30_perm_0, x = var_230_cast_fp16)[name = string("transpose_100")];
tensor<fp16, [1, 8, 192, 192]> matrix_ac_1_cast_fp16 = matmul(transpose_x = matrix_ac_1_transpose_x_0, transpose_y = matrix_ac_1_transpose_y_0, x = transpose_30, y = transpose_31)[name = string("matrix_ac_1_cast_fp16")];
bool matrix_bd_1_transpose_x_0 = const()[name = string("matrix_bd_1_transpose_x_0"), val = bool(false)];
bool matrix_bd_1_transpose_y_0 = const()[name = string("matrix_bd_1_transpose_y_0"), val = bool(false)];
tensor<fp16, [1, 8, 64, 383]> var_237_to_fp16 = const()[name = string("op_237_to_fp16"), val = tensor<fp16, [1, 8, 64, 383]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12500032)))];
tensor<fp16, [1, 8, 192, 64]> q_with_bias_v_1_cast_fp16 = transpose(perm = q_with_bias_v_1_perm_0, x = var_233_cast_fp16)[name = string("transpose_101")];
tensor<fp16, [1, 8, 192, 383]> matrix_bd_1_cast_fp16 = matmul(transpose_x = matrix_bd_1_transpose_x_0, transpose_y = matrix_bd_1_transpose_y_0, x = q_with_bias_v_1_cast_fp16, y = var_237_to_fp16)[name = string("matrix_bd_1_cast_fp16")];
bool x_padded_1_interleave_0 = const()[name = string("x_padded_1_interleave_0"), val = bool(false)];
tensor<fp16, [1, 8, 192, 1]> zero_pad_1_to_fp16 = const()[name = string("zero_pad_1_to_fp16"), val = tensor<fp16, [1, 8, 192, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12892288)))];
tensor<fp16, [1, 8, 192, 384]> x_padded_1_cast_fp16 = concat(axis = var_77, interleave = x_padded_1_interleave_0, values = (zero_pad_1_to_fp16, matrix_bd_1_cast_fp16))[name = string("x_padded_1_cast_fp16")];
tensor<int32, [4]> var_254 = const()[name = string("op_254"), val = tensor<int32, [4]>([1, 8, 384, 192])];
tensor<fp16, [1, 8, 384, 192]> x_padded_3_cast_fp16 = reshape(shape = var_254, x = x_padded_1_cast_fp16)[name = string("x_padded_3_cast_fp16")];
tensor<int32, [4]> var_258_begin_0 = const()[name = string("op_258_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
tensor<int32, [4]> var_258_end_0 = const()[name = string("op_258_end_0"), val = tensor<int32, [4]>([1, 8, 384, 192])];
tensor<bool, [4]> var_258_end_mask_0 = const()[name = string("op_258_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
tensor<fp16, [1, 8, 383, 192]> var_258_cast_fp16 = slice_by_index(begin = var_258_begin_0, end = var_258_end_0, end_mask = var_258_end_mask_0, x = x_padded_3_cast_fp16)[name = string("op_258_cast_fp16")];
tensor<int32, [4]> var_263 = const()[name = string("op_263"), val = tensor<int32, [4]>([1, 8, 192, 383])];
tensor<fp16, [1, 8, 192, 383]> shifted_1_cast_fp16 = reshape(shape = var_263, x = var_258_cast_fp16)[name = string("shifted_1_cast_fp16")];
tensor<int32, [4]> matrix_bd_3_begin_0 = const()[name = string("matrix_bd_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [4]> matrix_bd_3_end_0 = const()[name = string("matrix_bd_3_end_0"), val = tensor<int32, [4]>([1, 8, 192, 192])];
tensor<bool, [4]> matrix_bd_3_end_mask_0 = const()[name = string("matrix_bd_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
tensor<fp16, [1, 8, 192, 192]> matrix_bd_3_cast_fp16 = slice_by_index(begin = matrix_bd_3_begin_0, end = matrix_bd_3_end_0, end_mask = matrix_bd_3_end_mask_0, x = shifted_1_cast_fp16)[name = string("matrix_bd_3_cast_fp16")];
tensor<fp16, [1, 8, 192, 192]> var_275_cast_fp16 = add(x = matrix_ac_1_cast_fp16, y = matrix_bd_3_cast_fp16)[name = string("op_275_cast_fp16")];
fp16 _inversed_scores_1_y_0_to_fp16 = const()[name = string("_inversed_scores_1_y_0_to_fp16"), val = fp16(0x1p-3)];
tensor<fp16, [1, 8, 192, 192]> _inversed_scores_1_cast_fp16 = mul(x = var_275_cast_fp16, y = _inversed_scores_1_y_0_to_fp16)[name = string("_inversed_scores_1_cast_fp16")];
tensor<int32, [1]> var_279_axes_0 = const()[name = string("op_279_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, [1, 1, 1, 192]> var_279_cast_fp16 = expand_dims(axes = var_279_axes_0, x = mask_1_cast_fp16)[name = string("op_279_cast_fp16")];
fp16 var_79_promoted_1_to_fp16 = const()[name = string("op_79_promoted_1_to_fp16"), val = fp16(0x0p+0)];
tensor<bool, [1, 1, 1, 192]> mask_3_cast_fp16 = equal(x = var_279_cast_fp16, y = var_79_promoted_1_to_fp16)[name = string("mask_3_cast_fp16")];
fp16 var_59_to_fp16 = const()[name = string("op_59_to_fp16"), val = fp16(-inf)];
tensor<fp16, [1, 8, 192, 192]> scores_3_cast_fp16 = select(a = var_59_to_fp16, b = _inversed_scores_1_cast_fp16, cond = mask_3_cast_fp16)[name = string("scores_3_cast_fp16")];
tensor<fp16, [1, 8, 192, 192]> var_287_cast_fp16 = softmax(axis = var_77, x = scores_3_cast_fp16)[name = string("op_287_cast_fp16")];
fp16 var_68_to_fp16 = const()[name = string("op_68_to_fp16"), val = fp16(0x0p+0)];
tensor<fp16, [1, 8, 192, 192]> input_19_cast_fp16 = select(a = var_68_to_fp16, b = var_287_cast_fp16, cond = mask_3_cast_fp16)[name = string("input_19_cast_fp16")];
bool x_5_transpose_x_0 = const()[name = string("x_5_transpose_x_0"), val = bool(false)];
bool x_5_transpose_y_0 = const()[name = string("x_5_transpose_y_0"), val = bool(false)];
tensor<fp16, [1, 8, 192, 64]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = v_1_cast_fp16)[name = string("transpose_102")];
tensor<fp16, [1, 8, 192, 64]> x_5_cast_fp16 = matmul(transpose_x = x_5_transpose_x_0, transpose_y = x_5_transpose_y_0, x = input_19_cast_fp16, y = v_3_cast_fp16)[name = string("x_5_cast_fp16")];
tensor<int32, [4]> var_291_perm_0 = const()[name = string("op_291_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<int32, [3]> var_293 = const()[name = string("op_293"), val = tensor<int32, [3]>([1, -1, 512])];
tensor<fp16, [1, 192, 8, 64]> var_291_cast_fp16 = transpose(perm = var_291_perm_0, x = x_5_cast_fp16)[name = string("transpose_98")];
tensor<fp16, [1, 192, 512]> input_21_cast_fp16 = reshape(shape = var_293, x = var_291_cast_fp16)[name = string("input_21_cast_fp16")];
tensor<fp16, [512, 512]> encoder_encoders_0_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_encoders_0_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12895424)))];
tensor<fp16, [512]> encoder_encoders_0_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_encoders_0_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13419776)))];
tensor<fp16, [1, 192, 512]> linear_5_cast_fp16 = linear(bias = encoder_encoders_0_self_attn_linear_out_bias_to_fp16, weight = encoder_encoders_0_self_attn_linear_out_weight_to_fp16, x = input_21_cast_fp16)[name = string("linear_5_cast_fp16")];
tensor<fp16, [1, 192, 512]> input_25_cast_fp16 = add(x = input_17_cast_fp16, y = linear_5_cast_fp16)[name = string("input_25_cast_fp16")];
tensor<int32, [1]> input_27_axes_0 = const()[name = string("input_27_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [512]> encoder_encoders_0_norm_ff_weight_to_fp16 = const()[name = string("encoder_encoders_0_norm_ff_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13420864)))];
tensor<fp16, [512]> encoder_encoders_0_norm_ff_bias_to_fp16 = const()[name = string("encoder_encoders_0_norm_ff_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13421952)))];
tensor<fp16, [1, 192, 512]> input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = encoder_encoders_0_norm_ff_bias_to_fp16, epsilon = var_58_to_fp16, gamma = encoder_encoders_0_norm_ff_weight_to_fp16, x = input_25_cast_fp16)[name = string("input_27_cast_fp16")];
tensor<fp16, [2048, 512]> encoder_encoders_0_feed_forward_w_1_weight_to_fp16 = const()[name = string("encoder_encoders_0_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13423040)))];
tensor<fp16, [2048]> encoder_encoders_0_feed_forward_w_1_bias_to_fp16 = const()[name = string("encoder_encoders_0_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15520256)))];
tensor<fp16, [1, 192, 2048]> linear_6_cast_fp16 = linear(bias = encoder_encoders_0_feed_forward_w_1_bias_to_fp16, weight = encoder_encoders_0_feed_forward_w_1_weight_to_fp16, x = input_27_cast_fp16)[name = string("linear_6_cast_fp16")];
tensor<fp16, [1, 192, 2048]> input_31_cast_fp16 = silu(x = linear_6_cast_fp16)[name = string("input_31_cast_fp16")];
tensor<fp16, [512, 2048]> encoder_encoders_0_feed_forward_w_2_weight_to_fp16 = const()[name = string("encoder_encoders_0_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15524416)))];
tensor<fp16, [512]> encoder_encoders_0_feed_forward_w_2_bias_to_fp16 = const()[name = string("encoder_encoders_0_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17621632)))];
tensor<fp16, [1, 192, 512]> linear_7_cast_fp16 = linear(bias = encoder_encoders_0_feed_forward_w_2_bias_to_fp16, weight = encoder_encoders_0_feed_forward_w_2_weight_to_fp16, x = input_31_cast_fp16)[name = string("linear_7_cast_fp16")];
tensor<fp16, [1, 192, 512]> input_37_cast_fp16 = add(x = input_25_cast_fp16, y = linear_7_cast_fp16)[name = string("input_37_cast_fp16")];
tensor<int32, [1]> query_3_axes_0 = const()[name = string("query_3_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [512]> encoder_encoders_1_norm_mha_weight_to_fp16 = const()[name = string("encoder_encoders_1_norm_mha_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17622720)))];
tensor<fp16, [512]> encoder_encoders_1_norm_mha_bias_to_fp16 = const()[name = string("encoder_encoders_1_norm_mha_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17623808)))];
tensor<fp16, [1, 192, 512]> query_3_cast_fp16 = layer_norm(axes = query_3_axes_0, beta = encoder_encoders_1_norm_mha_bias_to_fp16, epsilon = var_58_to_fp16, gamma = encoder_encoders_1_norm_mha_weight_to_fp16, x = input_37_cast_fp16)[name = string("query_3_cast_fp16")];
tensor<fp16, [512, 512]> encoder_encoders_1_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_encoders_1_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17624896)))];
tensor<fp16, [512]> encoder_encoders_1_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_encoders_1_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18149248)))];
tensor<fp16, [1, 192, 512]> linear_8_cast_fp16 = linear(bias = encoder_encoders_1_self_attn_linear_q_bias_to_fp16, weight = encoder_encoders_1_self_attn_linear_q_weight_to_fp16, x = query_3_cast_fp16)[name = string("linear_8_cast_fp16")];
tensor<int32, [4]> var_343 = const()[name = string("op_343"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 192, 8, 64]> q_7_cast_fp16 = reshape(shape = var_343, x = linear_8_cast_fp16)[name = string("q_7_cast_fp16")];
tensor<fp16, [512, 512]> encoder_encoders_1_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_encoders_1_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18150336)))];
tensor<fp16, [512]> encoder_encoders_1_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_encoders_1_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18674688)))];
tensor<fp16, [1, 192, 512]> linear_9_cast_fp16 = linear(bias = encoder_encoders_1_self_attn_linear_k_bias_to_fp16, weight = encoder_encoders_1_self_attn_linear_k_weight_to_fp16, x = query_3_cast_fp16)[name = string("linear_9_cast_fp16")];
tensor<int32, [4]> var_348 = const()[name = string("op_348"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 192, 8, 64]> k_5_cast_fp16 = reshape(shape = var_348, x = linear_9_cast_fp16)[name = string("k_5_cast_fp16")];
tensor<fp16, [512, 512]> encoder_encoders_1_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_encoders_1_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18675776)))];
tensor<fp16, [512]> encoder_encoders_1_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_encoders_1_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19200128)))];
tensor<fp16, [1, 192, 512]> linear_10_cast_fp16 = linear(bias = encoder_encoders_1_self_attn_linear_v_bias_to_fp16, weight = encoder_encoders_1_self_attn_linear_v_weight_to_fp16, x = query_3_cast_fp16)[name = string("linear_10_cast_fp16")];
tensor<int32, [4]> var_353 = const()[name = string("op_353"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 192, 8, 64]> v_5_cast_fp16 = reshape(shape = var_353, x = linear_10_cast_fp16)[name = string("v_5_cast_fp16")];
tensor<int32, [4]> v_7_perm_0 = const()[name = string("v_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<fp16, [8, 64]> const_31_to_fp16 = const()[name = string("const_31_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19201216)))];
tensor<fp16, [1, 192, 8, 64]> var_366_cast_fp16 = add(x = q_7_cast_fp16, y = const_31_to_fp16)[name = string("op_366_cast_fp16")];
tensor<fp16, [8, 64]> const_32_to_fp16 = const()[name = string("const_32_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19202304)))];
tensor<fp16, [1, 192, 8, 64]> var_369_cast_fp16 = add(x = q_7_cast_fp16, y = const_32_to_fp16)[name = string("op_369_cast_fp16")];
tensor<int32, [4]> q_with_bias_v_3_perm_0 = const()[name = string("q_with_bias_v_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
bool matrix_ac_3_transpose_x_0 = const()[name = string("matrix_ac_3_transpose_x_0"), val = bool(false)];
bool matrix_ac_3_transpose_y_0 = const()[name = string("matrix_ac_3_transpose_y_0"), val = bool(false)];
tensor<int32, [4]> transpose_32_perm_0 = const()[name = string("transpose_32_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
tensor<int32, [4]> transpose_33_perm_0 = const()[name = string("transpose_33_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
tensor<fp16, [1, 8, 64, 192]> transpose_33 = transpose(perm = transpose_33_perm_0, x = k_5_cast_fp16)[name = string("transpose_94")];
tensor<fp16, [1, 8, 192, 64]> transpose_32 = transpose(perm = transpose_32_perm_0, x = var_366_cast_fp16)[name = string("transpose_95")];
tensor<fp16, [1, 8, 192, 192]> matrix_ac_3_cast_fp16 = matmul(transpose_x = matrix_ac_3_transpose_x_0, transpose_y = matrix_ac_3_transpose_y_0, x = transpose_32, y = transpose_33)[name = string("matrix_ac_3_cast_fp16")];
bool matrix_bd_5_transpose_x_0 = const()[name = string("matrix_bd_5_transpose_x_0"), val = bool(false)];
bool matrix_bd_5_transpose_y_0 = const()[name = string("matrix_bd_5_transpose_y_0"), val = bool(false)];
tensor<fp16, [1, 8, 64, 383]> var_373_to_fp16 = const()[name = string("op_373_to_fp16"), val = tensor<fp16, [1, 8, 64, 383]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19203392)))];
tensor<fp16, [1, 8, 192, 64]> q_with_bias_v_3_cast_fp16 = transpose(perm = q_with_bias_v_3_perm_0, x = var_369_cast_fp16)[name = string("transpose_96")];
tensor<fp16, [1, 8, 192, 383]> matrix_bd_5_cast_fp16 = matmul(transpose_x = matrix_bd_5_transpose_x_0, transpose_y = matrix_bd_5_transpose_y_0, x = q_with_bias_v_3_cast_fp16, y = var_373_to_fp16)[name = string("matrix_bd_5_cast_fp16")];
bool x_padded_5_interleave_0 = const()[name = string("x_padded_5_interleave_0"), val = bool(false)];
tensor<fp16, [1, 8, 192, 384]> x_padded_5_cast_fp16 = concat(axis = var_77, interleave = x_padded_5_interleave_0, values = (zero_pad_1_to_fp16, matrix_bd_5_cast_fp16))[name = string("x_padded_5_cast_fp16")];
tensor<int32, [4]> var_390 = const()[name = string("op_390"), val = tensor<int32, [4]>([1, 8, 384, 192])];
tensor<fp16, [1, 8, 384, 192]> x_padded_7_cast_fp16 = reshape(shape = var_390, x = x_padded_5_cast_fp16)[name = string("x_padded_7_cast_fp16")];
tensor<int32, [4]> var_394_begin_0 = const()[name = string("op_394_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
tensor<int32, [4]> var_394_end_0 = const()[name = string("op_394_end_0"), val = tensor<int32, [4]>([1, 8, 384, 192])];
tensor<bool, [4]> var_394_end_mask_0 = const()[name = string("op_394_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
tensor<fp16, [1, 8, 383, 192]> var_394_cast_fp16 = slice_by_index(begin = var_394_begin_0, end = var_394_end_0, end_mask = var_394_end_mask_0, x = x_padded_7_cast_fp16)[name = string("op_394_cast_fp16")];
tensor<int32, [4]> var_399 = const()[name = string("op_399"), val = tensor<int32, [4]>([1, 8, 192, 383])];
tensor<fp16, [1, 8, 192, 383]> shifted_3_cast_fp16 = reshape(shape = var_399, x = var_394_cast_fp16)[name = string("shifted_3_cast_fp16")];
tensor<int32, [4]> matrix_bd_7_begin_0 = const()[name = string("matrix_bd_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [4]> matrix_bd_7_end_0 = const()[name = string("matrix_bd_7_end_0"), val = tensor<int32, [4]>([1, 8, 192, 192])];
tensor<bool, [4]> matrix_bd_7_end_mask_0 = const()[name = string("matrix_bd_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
tensor<fp16, [1, 8, 192, 192]> matrix_bd_7_cast_fp16 = slice_by_index(begin = matrix_bd_7_begin_0, end = matrix_bd_7_end_0, end_mask = matrix_bd_7_end_mask_0, x = shifted_3_cast_fp16)[name = string("matrix_bd_7_cast_fp16")];
tensor<fp16, [1, 8, 192, 192]> var_411_cast_fp16 = add(x = matrix_ac_3_cast_fp16, y = matrix_bd_7_cast_fp16)[name = string("op_411_cast_fp16")];
fp16 _inversed_scores_5_y_0_to_fp16 = const()[name = string("_inversed_scores_5_y_0_to_fp16"), val = fp16(0x1p-3)];
tensor<fp16, [1, 8, 192, 192]> _inversed_scores_5_cast_fp16 = mul(x = var_411_cast_fp16, y = _inversed_scores_5_y_0_to_fp16)[name = string("_inversed_scores_5_cast_fp16")];
tensor<fp16, [1, 8, 192, 192]> scores_7_cast_fp16 = select(a = var_59_to_fp16, b = _inversed_scores_5_cast_fp16, cond = mask_3_cast_fp16)[name = string("scores_7_cast_fp16")];
tensor<fp16, [1, 8, 192, 192]> var_423_cast_fp16 = softmax(axis = var_77, x = scores_7_cast_fp16)[name = string("op_423_cast_fp16")];
tensor<fp16, [1, 8, 192, 192]> input_39_cast_fp16 = select(a = var_68_to_fp16, b = var_423_cast_fp16, cond = mask_3_cast_fp16)[name = string("input_39_cast_fp16")];
bool x_7_transpose_x_0 = const()[name = string("x_7_transpose_x_0"), val = bool(false)];
bool x_7_transpose_y_0 = const()[name = string("x_7_transpose_y_0"), val = bool(false)];
tensor<fp16, [1, 8, 192, 64]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = v_5_cast_fp16)[name = string("transpose_97")];
tensor<fp16, [1, 8, 192, 64]> x_7_cast_fp16 = matmul(transpose_x = x_7_transpose_x_0, transpose_y = x_7_transpose_y_0, x = input_39_cast_fp16, y = v_7_cast_fp16)[name = string("x_7_cast_fp16")];
tensor<int32, [4]> var_427_perm_0 = const()[name = string("op_427_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<int32, [3]> var_429 = const()[name = string("op_429"), val = tensor<int32, [3]>([1, -1, 512])];
tensor<fp16, [1, 192, 8, 64]> var_427_cast_fp16 = transpose(perm = var_427_perm_0, x = x_7_cast_fp16)[name = string("transpose_93")];
tensor<fp16, [1, 192, 512]> input_41_cast_fp16 = reshape(shape = var_429, x = var_427_cast_fp16)[name = string("input_41_cast_fp16")];
tensor<fp16, [512, 512]> encoder_encoders_1_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_encoders_1_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19595648)))];
tensor<fp16, [512]> encoder_encoders_1_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_encoders_1_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20120000)))];
tensor<fp16, [1, 192, 512]> linear_12_cast_fp16 = linear(bias = encoder_encoders_1_self_attn_linear_out_bias_to_fp16, weight = encoder_encoders_1_self_attn_linear_out_weight_to_fp16, x = input_41_cast_fp16)[name = string("linear_12_cast_fp16")];
tensor<fp16, [1, 192, 512]> input_45_cast_fp16 = add(x = input_37_cast_fp16, y = linear_12_cast_fp16)[name = string("input_45_cast_fp16")];
tensor<int32, [1]> input_47_axes_0 = const()[name = string("input_47_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [512]> encoder_encoders_1_norm_ff_weight_to_fp16 = const()[name = string("encoder_encoders_1_norm_ff_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20121088)))];
tensor<fp16, [512]> encoder_encoders_1_norm_ff_bias_to_fp16 = const()[name = string("encoder_encoders_1_norm_ff_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20122176)))];
tensor<fp16, [1, 192, 512]> input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = encoder_encoders_1_norm_ff_bias_to_fp16, epsilon = var_58_to_fp16, gamma = encoder_encoders_1_norm_ff_weight_to_fp16, x = input_45_cast_fp16)[name = string("input_47_cast_fp16")];
tensor<fp16, [2048, 512]> encoder_encoders_1_feed_forward_w_1_weight_to_fp16 = const()[name = string("encoder_encoders_1_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20123264)))];
tensor<fp16, [2048]> encoder_encoders_1_feed_forward_w_1_bias_to_fp16 = const()[name = string("encoder_encoders_1_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22220480)))];
tensor<fp16, [1, 192, 2048]> linear_13_cast_fp16 = linear(bias = encoder_encoders_1_feed_forward_w_1_bias_to_fp16, weight = encoder_encoders_1_feed_forward_w_1_weight_to_fp16, x = input_47_cast_fp16)[name = string("linear_13_cast_fp16")];
tensor<fp16, [1, 192, 2048]> input_51_cast_fp16 = silu(x = linear_13_cast_fp16)[name = string("input_51_cast_fp16")];
tensor<fp16, [512, 2048]> encoder_encoders_1_feed_forward_w_2_weight_to_fp16 = const()[name = string("encoder_encoders_1_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22224640)))];
tensor<fp16, [512]> encoder_encoders_1_feed_forward_w_2_bias_to_fp16 = const()[name = string("encoder_encoders_1_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24321856)))];
tensor<fp16, [1, 192, 512]> linear_14_cast_fp16 = linear(bias = encoder_encoders_1_feed_forward_w_2_bias_to_fp16, weight = encoder_encoders_1_feed_forward_w_2_weight_to_fp16, x = input_51_cast_fp16)[name = string("linear_14_cast_fp16")];
tensor<fp16, [1, 192, 512]> input_57_cast_fp16 = add(x = input_45_cast_fp16, y = linear_14_cast_fp16)[name = string("input_57_cast_fp16")];
tensor<int32, [1]> query_5_axes_0 = const()[name = string("query_5_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [512]> encoder_encoders_2_norm_mha_weight_to_fp16 = const()[name = string("encoder_encoders_2_norm_mha_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24322944)))];
tensor<fp16, [512]> encoder_encoders_2_norm_mha_bias_to_fp16 = const()[name = string("encoder_encoders_2_norm_mha_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24324032)))];
tensor<fp16, [1, 192, 512]> query_5_cast_fp16 = layer_norm(axes = query_5_axes_0, beta = encoder_encoders_2_norm_mha_bias_to_fp16, epsilon = var_58_to_fp16, gamma = encoder_encoders_2_norm_mha_weight_to_fp16, x = input_57_cast_fp16)[name = string("query_5_cast_fp16")];
tensor<fp16, [512, 512]> encoder_encoders_2_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_encoders_2_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24325120)))];
tensor<fp16, [512]> encoder_encoders_2_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_encoders_2_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24849472)))];
tensor<fp16, [1, 192, 512]> linear_15_cast_fp16 = linear(bias = encoder_encoders_2_self_attn_linear_q_bias_to_fp16, weight = encoder_encoders_2_self_attn_linear_q_weight_to_fp16, x = query_5_cast_fp16)[name = string("linear_15_cast_fp16")];
tensor<int32, [4]> var_473 = const()[name = string("op_473"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 192, 8, 64]> q_13_cast_fp16 = reshape(shape = var_473, x = linear_15_cast_fp16)[name = string("q_13_cast_fp16")];
tensor<fp16, [512, 512]> encoder_encoders_2_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_encoders_2_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24850560)))];
tensor<fp16, [512]> encoder_encoders_2_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_encoders_2_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25374912)))];
tensor<fp16, [1, 192, 512]> linear_16_cast_fp16 = linear(bias = encoder_encoders_2_self_attn_linear_k_bias_to_fp16, weight = encoder_encoders_2_self_attn_linear_k_weight_to_fp16, x = query_5_cast_fp16)[name = string("linear_16_cast_fp16")];
tensor<int32, [4]> var_478 = const()[name = string("op_478"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 192, 8, 64]> k_9_cast_fp16 = reshape(shape = var_478, x = linear_16_cast_fp16)[name = string("k_9_cast_fp16")];
tensor<fp16, [512, 512]> encoder_encoders_2_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_encoders_2_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25376000)))];
tensor<fp16, [512]> encoder_encoders_2_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_encoders_2_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25900352)))];
tensor<fp16, [1, 192, 512]> linear_17_cast_fp16 = linear(bias = encoder_encoders_2_self_attn_linear_v_bias_to_fp16, weight = encoder_encoders_2_self_attn_linear_v_weight_to_fp16, x = query_5_cast_fp16)[name = string("linear_17_cast_fp16")];
tensor<int32, [4]> var_483 = const()[name = string("op_483"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 192, 8, 64]> v_9_cast_fp16 = reshape(shape = var_483, x = linear_17_cast_fp16)[name = string("v_9_cast_fp16")];
tensor<int32, [4]> v_11_perm_0 = const()[name = string("v_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<fp16, [8, 64]> const_49_to_fp16 = const()[name = string("const_49_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25901440)))];
tensor<fp16, [1, 192, 8, 64]> var_496_cast_fp16 = add(x = q_13_cast_fp16, y = const_49_to_fp16)[name = string("op_496_cast_fp16")];
tensor<fp16, [8, 64]> const_50_to_fp16 = const()[name = string("const_50_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25902528)))];
tensor<fp16, [1, 192, 8, 64]> var_499_cast_fp16 = add(x = q_13_cast_fp16, y = const_50_to_fp16)[name = string("op_499_cast_fp16")];
tensor<int32, [4]> q_with_bias_v_5_perm_0 = const()[name = string("q_with_bias_v_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
bool matrix_ac_5_transpose_x_0 = const()[name = string("matrix_ac_5_transpose_x_0"), val = bool(false)];
bool matrix_ac_5_transpose_y_0 = const()[name = string("matrix_ac_5_transpose_y_0"), val = bool(false)];
tensor<int32, [4]> transpose_34_perm_0 = const()[name = string("transpose_34_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
tensor<int32, [4]> transpose_35_perm_0 = const()[name = string("transpose_35_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
tensor<fp16, [1, 8, 64, 192]> transpose_35 = transpose(perm = transpose_35_perm_0, x = k_9_cast_fp16)[name = string("transpose_89")];
tensor<fp16, [1, 8, 192, 64]> transpose_34 = transpose(perm = transpose_34_perm_0, x = var_496_cast_fp16)[name = string("transpose_90")];
tensor<fp16, [1, 8, 192, 192]> matrix_ac_5_cast_fp16 = matmul(transpose_x = matrix_ac_5_transpose_x_0, transpose_y = matrix_ac_5_transpose_y_0, x = transpose_34, y = transpose_35)[name = string("matrix_ac_5_cast_fp16")];
bool matrix_bd_9_transpose_x_0 = const()[name = string("matrix_bd_9_transpose_x_0"), val = bool(false)];
bool matrix_bd_9_transpose_y_0 = const()[name = string("matrix_bd_9_transpose_y_0"), val = bool(false)];
tensor<fp16, [1, 8, 64, 383]> var_503_to_fp16 = const()[name = string("op_503_to_fp16"), val = tensor<fp16, [1, 8, 64, 383]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25903616)))];
tensor<fp16, [1, 8, 192, 64]> q_with_bias_v_5_cast_fp16 = transpose(perm = q_with_bias_v_5_perm_0, x = var_499_cast_fp16)[name = string("transpose_91")];
tensor<fp16, [1, 8, 192, 383]> matrix_bd_9_cast_fp16 = matmul(transpose_x = matrix_bd_9_transpose_x_0, transpose_y = matrix_bd_9_transpose_y_0, x = q_with_bias_v_5_cast_fp16, y = var_503_to_fp16)[name = string("matrix_bd_9_cast_fp16")];
bool x_padded_9_interleave_0 = const()[name = string("x_padded_9_interleave_0"), val = bool(false)];
tensor<fp16, [1, 8, 192, 384]> x_padded_9_cast_fp16 = concat(axis = var_77, interleave = x_padded_9_interleave_0, values = (zero_pad_1_to_fp16, matrix_bd_9_cast_fp16))[name = string("x_padded_9_cast_fp16")];
tensor<int32, [4]> var_520 = const()[name = string("op_520"), val = tensor<int32, [4]>([1, 8, 384, 192])];
tensor<fp16, [1, 8, 384, 192]> x_padded_11_cast_fp16 = reshape(shape = var_520, x = x_padded_9_cast_fp16)[name = string("x_padded_11_cast_fp16")];
tensor<int32, [4]> var_524_begin_0 = const()[name = string("op_524_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
tensor<int32, [4]> var_524_end_0 = const()[name = string("op_524_end_0"), val = tensor<int32, [4]>([1, 8, 384, 192])];
tensor<bool, [4]> var_524_end_mask_0 = const()[name = string("op_524_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
tensor<fp16, [1, 8, 383, 192]> var_524_cast_fp16 = slice_by_index(begin = var_524_begin_0, end = var_524_end_0, end_mask = var_524_end_mask_0, x = x_padded_11_cast_fp16)[name = string("op_524_cast_fp16")];
tensor<int32, [4]> var_529 = const()[name = string("op_529"), val = tensor<int32, [4]>([1, 8, 192, 383])];
tensor<fp16, [1, 8, 192, 383]> shifted_5_cast_fp16 = reshape(shape = var_529, x = var_524_cast_fp16)[name = string("shifted_5_cast_fp16")];
tensor<int32, [4]> matrix_bd_11_begin_0 = const()[name = string("matrix_bd_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [4]> matrix_bd_11_end_0 = const()[name = string("matrix_bd_11_end_0"), val = tensor<int32, [4]>([1, 8, 192, 192])];
tensor<bool, [4]> matrix_bd_11_end_mask_0 = const()[name = string("matrix_bd_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
tensor<fp16, [1, 8, 192, 192]> matrix_bd_11_cast_fp16 = slice_by_index(begin = matrix_bd_11_begin_0, end = matrix_bd_11_end_0, end_mask = matrix_bd_11_end_mask_0, x = shifted_5_cast_fp16)[name = string("matrix_bd_11_cast_fp16")];
tensor<fp16, [1, 8, 192, 192]> var_541_cast_fp16 = add(x = matrix_ac_5_cast_fp16, y = matrix_bd_11_cast_fp16)[name = string("op_541_cast_fp16")];
fp16 _inversed_scores_9_y_0_to_fp16 = const()[name = string("_inversed_scores_9_y_0_to_fp16"), val = fp16(0x1p-3)];
tensor<fp16, [1, 8, 192, 192]> _inversed_scores_9_cast_fp16 = mul(x = var_541_cast_fp16, y = _inversed_scores_9_y_0_to_fp16)[name = string("_inversed_scores_9_cast_fp16")];
tensor<fp16, [1, 8, 192, 192]> scores_11_cast_fp16 = select(a = var_59_to_fp16, b = _inversed_scores_9_cast_fp16, cond = mask_3_cast_fp16)[name = string("scores_11_cast_fp16")];
tensor<fp16, [1, 8, 192, 192]> var_553_cast_fp16 = softmax(axis = var_77, x = scores_11_cast_fp16)[name = string("op_553_cast_fp16")];
tensor<fp16, [1, 8, 192, 192]> input_59_cast_fp16 = select(a = var_68_to_fp16, b = var_553_cast_fp16, cond = mask_3_cast_fp16)[name = string("input_59_cast_fp16")];
bool x_9_transpose_x_0 = const()[name = string("x_9_transpose_x_0"), val = bool(false)];
bool x_9_transpose_y_0 = const()[name = string("x_9_transpose_y_0"), val = bool(false)];
tensor<fp16, [1, 8, 192, 64]> v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = v_9_cast_fp16)[name = string("transpose_92")];
tensor<fp16, [1, 8, 192, 64]> x_9_cast_fp16 = matmul(transpose_x = x_9_transpose_x_0, transpose_y = x_9_transpose_y_0, x = input_59_cast_fp16, y = v_11_cast_fp16)[name = string("x_9_cast_fp16")];
tensor<int32, [4]> var_557_perm_0 = const()[name = string("op_557_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<int32, [3]> var_559 = const()[name = string("op_559"), val = tensor<int32, [3]>([1, -1, 512])];
tensor<fp16, [1, 192, 8, 64]> var_557_cast_fp16 = transpose(perm = var_557_perm_0, x = x_9_cast_fp16)[name = string("transpose_88")];
tensor<fp16, [1, 192, 512]> input_61_cast_fp16 = reshape(shape = var_559, x = var_557_cast_fp16)[name = string("input_61_cast_fp16")];
tensor<fp16, [512, 512]> encoder_encoders_2_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_encoders_2_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26295872)))];
tensor<fp16, [512]> encoder_encoders_2_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_encoders_2_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26820224)))];
tensor<fp16, [1, 192, 512]> linear_19_cast_fp16 = linear(bias = encoder_encoders_2_self_attn_linear_out_bias_to_fp16, weight = encoder_encoders_2_self_attn_linear_out_weight_to_fp16, x = input_61_cast_fp16)[name = string("linear_19_cast_fp16")];
tensor<fp16, [1, 192, 512]> input_65_cast_fp16 = add(x = input_57_cast_fp16, y = linear_19_cast_fp16)[name = string("input_65_cast_fp16")];
tensor<int32, [1]> input_67_axes_0 = const()[name = string("input_67_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [512]> encoder_encoders_2_norm_ff_weight_to_fp16 = const()[name = string("encoder_encoders_2_norm_ff_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26821312)))];
tensor<fp16, [512]> encoder_encoders_2_norm_ff_bias_to_fp16 = const()[name = string("encoder_encoders_2_norm_ff_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26822400)))];
tensor<fp16, [1, 192, 512]> input_67_cast_fp16 = layer_norm(axes = input_67_axes_0, beta = encoder_encoders_2_norm_ff_bias_to_fp16, epsilon = var_58_to_fp16, gamma = encoder_encoders_2_norm_ff_weight_to_fp16, x = input_65_cast_fp16)[name = string("input_67_cast_fp16")];
tensor<fp16, [2048, 512]> encoder_encoders_2_feed_forward_w_1_weight_to_fp16 = const()[name = string("encoder_encoders_2_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26823488)))];
tensor<fp16, [2048]> encoder_encoders_2_feed_forward_w_1_bias_to_fp16 = const()[name = string("encoder_encoders_2_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28920704)))];
tensor<fp16, [1, 192, 2048]> linear_20_cast_fp16 = linear(bias = encoder_encoders_2_feed_forward_w_1_bias_to_fp16, weight = encoder_encoders_2_feed_forward_w_1_weight_to_fp16, x = input_67_cast_fp16)[name = string("linear_20_cast_fp16")];
tensor<fp16, [1, 192, 2048]> input_71_cast_fp16 = silu(x = linear_20_cast_fp16)[name = string("input_71_cast_fp16")];
tensor<fp16, [512, 2048]> encoder_encoders_2_feed_forward_w_2_weight_to_fp16 = const()[name = string("encoder_encoders_2_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28924864)))];
tensor<fp16, [512]> encoder_encoders_2_feed_forward_w_2_bias_to_fp16 = const()[name = string("encoder_encoders_2_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31022080)))];
tensor<fp16, [1, 192, 512]> linear_21_cast_fp16 = linear(bias = encoder_encoders_2_feed_forward_w_2_bias_to_fp16, weight = encoder_encoders_2_feed_forward_w_2_weight_to_fp16, x = input_71_cast_fp16)[name = string("linear_21_cast_fp16")];
tensor<fp16, [1, 192, 512]> input_77_cast_fp16 = add(x = input_65_cast_fp16, y = linear_21_cast_fp16)[name = string("input_77_cast_fp16")];
tensor<int32, [1]> query_7_axes_0 = const()[name = string("query_7_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [512]> encoder_encoders_3_norm_mha_weight_to_fp16 = const()[name = string("encoder_encoders_3_norm_mha_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31023168)))];
tensor<fp16, [512]> encoder_encoders_3_norm_mha_bias_to_fp16 = const()[name = string("encoder_encoders_3_norm_mha_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31024256)))];
tensor<fp16, [1, 192, 512]> query_7_cast_fp16 = layer_norm(axes = query_7_axes_0, beta = encoder_encoders_3_norm_mha_bias_to_fp16, epsilon = var_58_to_fp16, gamma = encoder_encoders_3_norm_mha_weight_to_fp16, x = input_77_cast_fp16)[name = string("query_7_cast_fp16")];
tensor<fp16, [512, 512]> encoder_encoders_3_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_encoders_3_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31025344)))];
tensor<fp16, [512]> encoder_encoders_3_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_encoders_3_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31549696)))];
tensor<fp16, [1, 192, 512]> linear_22_cast_fp16 = linear(bias = encoder_encoders_3_self_attn_linear_q_bias_to_fp16, weight = encoder_encoders_3_self_attn_linear_q_weight_to_fp16, x = query_7_cast_fp16)[name = string("linear_22_cast_fp16")];
tensor<int32, [4]> var_603 = const()[name = string("op_603"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 192, 8, 64]> q_19_cast_fp16 = reshape(shape = var_603, x = linear_22_cast_fp16)[name = string("q_19_cast_fp16")];
tensor<fp16, [512, 512]> encoder_encoders_3_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_encoders_3_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31550784)))];
tensor<fp16, [512]> encoder_encoders_3_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_encoders_3_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32075136)))];
tensor<fp16, [1, 192, 512]> linear_23_cast_fp16 = linear(bias = encoder_encoders_3_self_attn_linear_k_bias_to_fp16, weight = encoder_encoders_3_self_attn_linear_k_weight_to_fp16, x = query_7_cast_fp16)[name = string("linear_23_cast_fp16")];
tensor<int32, [4]> var_608 = const()[name = string("op_608"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 192, 8, 64]> k_13_cast_fp16 = reshape(shape = var_608, x = linear_23_cast_fp16)[name = string("k_13_cast_fp16")];
tensor<fp16, [512, 512]> encoder_encoders_3_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_encoders_3_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32076224)))];
tensor<fp16, [512]> encoder_encoders_3_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_encoders_3_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32600576)))];
tensor<fp16, [1, 192, 512]> linear_24_cast_fp16 = linear(bias = encoder_encoders_3_self_attn_linear_v_bias_to_fp16, weight = encoder_encoders_3_self_attn_linear_v_weight_to_fp16, x = query_7_cast_fp16)[name = string("linear_24_cast_fp16")];
tensor<int32, [4]> var_613 = const()[name = string("op_613"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 192, 8, 64]> v_13_cast_fp16 = reshape(shape = var_613, x = linear_24_cast_fp16)[name = string("v_13_cast_fp16")];
tensor<int32, [4]> v_15_perm_0 = const()[name = string("v_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<fp16, [8, 64]> const_67_to_fp16 = const()[name = string("const_67_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32601664)))];
tensor<fp16, [1, 192, 8, 64]> var_626_cast_fp16 = add(x = q_19_cast_fp16, y = const_67_to_fp16)[name = string("op_626_cast_fp16")];
tensor<fp16, [8, 64]> const_68_to_fp16 = const()[name = string("const_68_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32602752)))];
tensor<fp16, [1, 192, 8, 64]> var_629_cast_fp16 = add(x = q_19_cast_fp16, y = const_68_to_fp16)[name = string("op_629_cast_fp16")];
tensor<int32, [4]> q_with_bias_v_7_perm_0 = const()[name = string("q_with_bias_v_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
bool matrix_ac_7_transpose_x_0 = const()[name = string("matrix_ac_7_transpose_x_0"), val = bool(false)];
bool matrix_ac_7_transpose_y_0 = const()[name = string("matrix_ac_7_transpose_y_0"), val = bool(false)];
tensor<int32, [4]> transpose_36_perm_0 = const()[name = string("transpose_36_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
tensor<int32, [4]> transpose_37_perm_0 = const()[name = string("transpose_37_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
tensor<fp16, [1, 8, 64, 192]> transpose_37 = transpose(perm = transpose_37_perm_0, x = k_13_cast_fp16)[name = string("transpose_84")];
tensor<fp16, [1, 8, 192, 64]> transpose_36 = transpose(perm = transpose_36_perm_0, x = var_626_cast_fp16)[name = string("transpose_85")];
tensor<fp16, [1, 8, 192, 192]> matrix_ac_7_cast_fp16 = matmul(transpose_x = matrix_ac_7_transpose_x_0, transpose_y = matrix_ac_7_transpose_y_0, x = transpose_36, y = transpose_37)[name = string("matrix_ac_7_cast_fp16")];
bool matrix_bd_13_transpose_x_0 = const()[name = string("matrix_bd_13_transpose_x_0"), val = bool(false)];
bool matrix_bd_13_transpose_y_0 = const()[name = string("matrix_bd_13_transpose_y_0"), val = bool(false)];
tensor<fp16, [1, 8, 64, 383]> var_633_to_fp16 = const()[name = string("op_633_to_fp16"), val = tensor<fp16, [1, 8, 64, 383]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32603840)))];
tensor<fp16, [1, 8, 192, 64]> q_with_bias_v_7_cast_fp16 = transpose(perm = q_with_bias_v_7_perm_0, x = var_629_cast_fp16)[name = string("transpose_86")];
tensor<fp16, [1, 8, 192, 383]> matrix_bd_13_cast_fp16 = matmul(transpose_x = matrix_bd_13_transpose_x_0, transpose_y = matrix_bd_13_transpose_y_0, x = q_with_bias_v_7_cast_fp16, y = var_633_to_fp16)[name = string("matrix_bd_13_cast_fp16")];
bool x_padded_13_interleave_0 = const()[name = string("x_padded_13_interleave_0"), val = bool(false)];
tensor<fp16, [1, 8, 192, 384]> x_padded_13_cast_fp16 = concat(axis = var_77, interleave = x_padded_13_interleave_0, values = (zero_pad_1_to_fp16, matrix_bd_13_cast_fp16))[name = string("x_padded_13_cast_fp16")];
tensor<int32, [4]> var_650 = const()[name = string("op_650"), val = tensor<int32, [4]>([1, 8, 384, 192])];
tensor<fp16, [1, 8, 384, 192]> x_padded_15_cast_fp16 = reshape(shape = var_650, x = x_padded_13_cast_fp16)[name = string("x_padded_15_cast_fp16")];
tensor<int32, [4]> var_654_begin_0 = const()[name = string("op_654_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
tensor<int32, [4]> var_654_end_0 = const()[name = string("op_654_end_0"), val = tensor<int32, [4]>([1, 8, 384, 192])];
tensor<bool, [4]> var_654_end_mask_0 = const()[name = string("op_654_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
tensor<fp16, [1, 8, 383, 192]> var_654_cast_fp16 = slice_by_index(begin = var_654_begin_0, end = var_654_end_0, end_mask = var_654_end_mask_0, x = x_padded_15_cast_fp16)[name = string("op_654_cast_fp16")];
tensor<int32, [4]> var_659 = const()[name = string("op_659"), val = tensor<int32, [4]>([1, 8, 192, 383])];
tensor<fp16, [1, 8, 192, 383]> shifted_7_cast_fp16 = reshape(shape = var_659, x = var_654_cast_fp16)[name = string("shifted_7_cast_fp16")];
tensor<int32, [4]> matrix_bd_15_begin_0 = const()[name = string("matrix_bd_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [4]> matrix_bd_15_end_0 = const()[name = string("matrix_bd_15_end_0"), val = tensor<int32, [4]>([1, 8, 192, 192])];
tensor<bool, [4]> matrix_bd_15_end_mask_0 = const()[name = string("matrix_bd_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
tensor<fp16, [1, 8, 192, 192]> matrix_bd_15_cast_fp16 = slice_by_index(begin = matrix_bd_15_begin_0, end = matrix_bd_15_end_0, end_mask = matrix_bd_15_end_mask_0, x = shifted_7_cast_fp16)[name = string("matrix_bd_15_cast_fp16")];
tensor<fp16, [1, 8, 192, 192]> var_671_cast_fp16 = add(x = matrix_ac_7_cast_fp16, y = matrix_bd_15_cast_fp16)[name = string("op_671_cast_fp16")];
fp16 _inversed_scores_13_y_0_to_fp16 = const()[name = string("_inversed_scores_13_y_0_to_fp16"), val = fp16(0x1p-3)];
tensor<fp16, [1, 8, 192, 192]> _inversed_scores_13_cast_fp16 = mul(x = var_671_cast_fp16, y = _inversed_scores_13_y_0_to_fp16)[name = string("_inversed_scores_13_cast_fp16")];
tensor<fp16, [1, 8, 192, 192]> scores_15_cast_fp16 = select(a = var_59_to_fp16, b = _inversed_scores_13_cast_fp16, cond = mask_3_cast_fp16)[name = string("scores_15_cast_fp16")];
tensor<fp16, [1, 8, 192, 192]> var_683_cast_fp16 = softmax(axis = var_77, x = scores_15_cast_fp16)[name = string("op_683_cast_fp16")];
tensor<fp16, [1, 8, 192, 192]> input_79_cast_fp16 = select(a = var_68_to_fp16, b = var_683_cast_fp16, cond = mask_3_cast_fp16)[name = string("input_79_cast_fp16")];
bool x_11_transpose_x_0 = const()[name = string("x_11_transpose_x_0"), val = bool(false)];
bool x_11_transpose_y_0 = const()[name = string("x_11_transpose_y_0"), val = bool(false)];
tensor<fp16, [1, 8, 192, 64]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = v_13_cast_fp16)[name = string("transpose_87")];
tensor<fp16, [1, 8, 192, 64]> x_11_cast_fp16 = matmul(transpose_x = x_11_transpose_x_0, transpose_y = x_11_transpose_y_0, x = input_79_cast_fp16, y = v_15_cast_fp16)[name = string("x_11_cast_fp16")];
tensor<int32, [4]> var_687_perm_0 = const()[name = string("op_687_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<int32, [3]> var_689 = const()[name = string("op_689"), val = tensor<int32, [3]>([1, -1, 512])];
tensor<fp16, [1, 192, 8, 64]> var_687_cast_fp16 = transpose(perm = var_687_perm_0, x = x_11_cast_fp16)[name = string("transpose_83")];
tensor<fp16, [1, 192, 512]> input_81_cast_fp16 = reshape(shape = var_689, x = var_687_cast_fp16)[name = string("input_81_cast_fp16")];
tensor<fp16, [512, 512]> encoder_encoders_3_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_encoders_3_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32996096)))];
tensor<fp16, [512]> encoder_encoders_3_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_encoders_3_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33520448)))];
tensor<fp16, [1, 192, 512]> linear_26_cast_fp16 = linear(bias = encoder_encoders_3_self_attn_linear_out_bias_to_fp16, weight = encoder_encoders_3_self_attn_linear_out_weight_to_fp16, x = input_81_cast_fp16)[name = string("linear_26_cast_fp16")];
tensor<fp16, [1, 192, 512]> input_85_cast_fp16 = add(x = input_77_cast_fp16, y = linear_26_cast_fp16)[name = string("input_85_cast_fp16")];
tensor<int32, [1]> input_87_axes_0 = const()[name = string("input_87_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [512]> encoder_encoders_3_norm_ff_weight_to_fp16 = const()[name = string("encoder_encoders_3_norm_ff_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33521536)))];
tensor<fp16, [512]> encoder_encoders_3_norm_ff_bias_to_fp16 = const()[name = string("encoder_encoders_3_norm_ff_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33522624)))];
tensor<fp16, [1, 192, 512]> input_87_cast_fp16 = layer_norm(axes = input_87_axes_0, beta = encoder_encoders_3_norm_ff_bias_to_fp16, epsilon = var_58_to_fp16, gamma = encoder_encoders_3_norm_ff_weight_to_fp16, x = input_85_cast_fp16)[name = string("input_87_cast_fp16")];
tensor<fp16, [2048, 512]> encoder_encoders_3_feed_forward_w_1_weight_to_fp16 = const()[name = string("encoder_encoders_3_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33523712)))];
tensor<fp16, [2048]> encoder_encoders_3_feed_forward_w_1_bias_to_fp16 = const()[name = string("encoder_encoders_3_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35620928)))];
tensor<fp16, [1, 192, 2048]> linear_27_cast_fp16 = linear(bias = encoder_encoders_3_feed_forward_w_1_bias_to_fp16, weight = encoder_encoders_3_feed_forward_w_1_weight_to_fp16, x = input_87_cast_fp16)[name = string("linear_27_cast_fp16")];
tensor<fp16, [1, 192, 2048]> input_91_cast_fp16 = silu(x = linear_27_cast_fp16)[name = string("input_91_cast_fp16")];
tensor<fp16, [512, 2048]> encoder_encoders_3_feed_forward_w_2_weight_to_fp16 = const()[name = string("encoder_encoders_3_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35625088)))];
tensor<fp16, [512]> encoder_encoders_3_feed_forward_w_2_bias_to_fp16 = const()[name = string("encoder_encoders_3_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37722304)))];
tensor<fp16, [1, 192, 512]> linear_28_cast_fp16 = linear(bias = encoder_encoders_3_feed_forward_w_2_bias_to_fp16, weight = encoder_encoders_3_feed_forward_w_2_weight_to_fp16, x = input_91_cast_fp16)[name = string("linear_28_cast_fp16")];
tensor<fp16, [1, 192, 512]> input_97_cast_fp16 = add(x = input_85_cast_fp16, y = linear_28_cast_fp16)[name = string("input_97_cast_fp16")];
tensor<int32, [1]> query_9_axes_0 = const()[name = string("query_9_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [512]> encoder_encoders_4_norm_mha_weight_to_fp16 = const()[name = string("encoder_encoders_4_norm_mha_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37723392)))];
tensor<fp16, [512]> encoder_encoders_4_norm_mha_bias_to_fp16 = const()[name = string("encoder_encoders_4_norm_mha_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37724480)))];
tensor<fp16, [1, 192, 512]> query_9_cast_fp16 = layer_norm(axes = query_9_axes_0, beta = encoder_encoders_4_norm_mha_bias_to_fp16, epsilon = var_58_to_fp16, gamma = encoder_encoders_4_norm_mha_weight_to_fp16, x = input_97_cast_fp16)[name = string("query_9_cast_fp16")];
tensor<fp16, [512, 512]> encoder_encoders_4_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_encoders_4_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37725568)))];
tensor<fp16, [512]> encoder_encoders_4_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_encoders_4_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38249920)))];
tensor<fp16, [1, 192, 512]> linear_29_cast_fp16 = linear(bias = encoder_encoders_4_self_attn_linear_q_bias_to_fp16, weight = encoder_encoders_4_self_attn_linear_q_weight_to_fp16, x = query_9_cast_fp16)[name = string("linear_29_cast_fp16")];
tensor<int32, [4]> var_733 = const()[name = string("op_733"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 192, 8, 64]> q_25_cast_fp16 = reshape(shape = var_733, x = linear_29_cast_fp16)[name = string("q_25_cast_fp16")];
tensor<fp16, [512, 512]> encoder_encoders_4_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_encoders_4_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38251008)))];
tensor<fp16, [512]> encoder_encoders_4_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_encoders_4_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38775360)))];
tensor<fp16, [1, 192, 512]> linear_30_cast_fp16 = linear(bias = encoder_encoders_4_self_attn_linear_k_bias_to_fp16, weight = encoder_encoders_4_self_attn_linear_k_weight_to_fp16, x = query_9_cast_fp16)[name = string("linear_30_cast_fp16")];
tensor<int32, [4]> var_738 = const()[name = string("op_738"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 192, 8, 64]> k_17_cast_fp16 = reshape(shape = var_738, x = linear_30_cast_fp16)[name = string("k_17_cast_fp16")];
tensor<fp16, [512, 512]> encoder_encoders_4_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_encoders_4_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38776448)))];
tensor<fp16, [512]> encoder_encoders_4_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_encoders_4_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39300800)))];
tensor<fp16, [1, 192, 512]> linear_31_cast_fp16 = linear(bias = encoder_encoders_4_self_attn_linear_v_bias_to_fp16, weight = encoder_encoders_4_self_attn_linear_v_weight_to_fp16, x = query_9_cast_fp16)[name = string("linear_31_cast_fp16")];
tensor<int32, [4]> var_743 = const()[name = string("op_743"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 192, 8, 64]> v_17_cast_fp16 = reshape(shape = var_743, x = linear_31_cast_fp16)[name = string("v_17_cast_fp16")];
tensor<int32, [4]> v_19_perm_0 = const()[name = string("v_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<fp16, [8, 64]> const_85_to_fp16 = const()[name = string("const_85_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39301888)))];
tensor<fp16, [1, 192, 8, 64]> var_756_cast_fp16 = add(x = q_25_cast_fp16, y = const_85_to_fp16)[name = string("op_756_cast_fp16")];
tensor<fp16, [8, 64]> const_86_to_fp16 = const()[name = string("const_86_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39302976)))];
tensor<fp16, [1, 192, 8, 64]> var_759_cast_fp16 = add(x = q_25_cast_fp16, y = const_86_to_fp16)[name = string("op_759_cast_fp16")];
tensor<int32, [4]> q_with_bias_v_9_perm_0 = const()[name = string("q_with_bias_v_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
bool matrix_ac_9_transpose_x_0 = const()[name = string("matrix_ac_9_transpose_x_0"), val = bool(false)];
bool matrix_ac_9_transpose_y_0 = const()[name = string("matrix_ac_9_transpose_y_0"), val = bool(false)];
tensor<int32, [4]> transpose_38_perm_0 = const()[name = string("transpose_38_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
tensor<int32, [4]> transpose_39_perm_0 = const()[name = string("transpose_39_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
tensor<fp16, [1, 8, 64, 192]> transpose_39 = transpose(perm = transpose_39_perm_0, x = k_17_cast_fp16)[name = string("transpose_79")];
tensor<fp16, [1, 8, 192, 64]> transpose_38 = transpose(perm = transpose_38_perm_0, x = var_756_cast_fp16)[name = string("transpose_80")];
tensor<fp16, [1, 8, 192, 192]> matrix_ac_9_cast_fp16 = matmul(transpose_x = matrix_ac_9_transpose_x_0, transpose_y = matrix_ac_9_transpose_y_0, x = transpose_38, y = transpose_39)[name = string("matrix_ac_9_cast_fp16")];
bool matrix_bd_17_transpose_x_0 = const()[name = string("matrix_bd_17_transpose_x_0"), val = bool(false)];
bool matrix_bd_17_transpose_y_0 = const()[name = string("matrix_bd_17_transpose_y_0"), val = bool(false)];
tensor<fp16, [1, 8, 64, 383]> var_763_to_fp16 = const()[name = string("op_763_to_fp16"), val = tensor<fp16, [1, 8, 64, 383]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39304064)))];
tensor<fp16, [1, 8, 192, 64]> q_with_bias_v_9_cast_fp16 = transpose(perm = q_with_bias_v_9_perm_0, x = var_759_cast_fp16)[name = string("transpose_81")];
tensor<fp16, [1, 8, 192, 383]> matrix_bd_17_cast_fp16 = matmul(transpose_x = matrix_bd_17_transpose_x_0, transpose_y = matrix_bd_17_transpose_y_0, x = q_with_bias_v_9_cast_fp16, y = var_763_to_fp16)[name = string("matrix_bd_17_cast_fp16")];
bool x_padded_17_interleave_0 = const()[name = string("x_padded_17_interleave_0"), val = bool(false)];
tensor<fp16, [1, 8, 192, 384]> x_padded_17_cast_fp16 = concat(axis = var_77, interleave = x_padded_17_interleave_0, values = (zero_pad_1_to_fp16, matrix_bd_17_cast_fp16))[name = string("x_padded_17_cast_fp16")];
tensor<int32, [4]> var_780 = const()[name = string("op_780"), val = tensor<int32, [4]>([1, 8, 384, 192])];
tensor<fp16, [1, 8, 384, 192]> x_padded_19_cast_fp16 = reshape(shape = var_780, x = x_padded_17_cast_fp16)[name = string("x_padded_19_cast_fp16")];
tensor<int32, [4]> var_784_begin_0 = const()[name = string("op_784_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
tensor<int32, [4]> var_784_end_0 = const()[name = string("op_784_end_0"), val = tensor<int32, [4]>([1, 8, 384, 192])];
tensor<bool, [4]> var_784_end_mask_0 = const()[name = string("op_784_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
tensor<fp16, [1, 8, 383, 192]> var_784_cast_fp16 = slice_by_index(begin = var_784_begin_0, end = var_784_end_0, end_mask = var_784_end_mask_0, x = x_padded_19_cast_fp16)[name = string("op_784_cast_fp16")];
tensor<int32, [4]> var_789 = const()[name = string("op_789"), val = tensor<int32, [4]>([1, 8, 192, 383])];
tensor<fp16, [1, 8, 192, 383]> shifted_9_cast_fp16 = reshape(shape = var_789, x = var_784_cast_fp16)[name = string("shifted_9_cast_fp16")];
tensor<int32, [4]> matrix_bd_19_begin_0 = const()[name = string("matrix_bd_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [4]> matrix_bd_19_end_0 = const()[name = string("matrix_bd_19_end_0"), val = tensor<int32, [4]>([1, 8, 192, 192])];
tensor<bool, [4]> matrix_bd_19_end_mask_0 = const()[name = string("matrix_bd_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
tensor<fp16, [1, 8, 192, 192]> matrix_bd_19_cast_fp16 = slice_by_index(begin = matrix_bd_19_begin_0, end = matrix_bd_19_end_0, end_mask = matrix_bd_19_end_mask_0, x = shifted_9_cast_fp16)[name = string("matrix_bd_19_cast_fp16")];
tensor<fp16, [1, 8, 192, 192]> var_801_cast_fp16 = add(x = matrix_ac_9_cast_fp16, y = matrix_bd_19_cast_fp16)[name = string("op_801_cast_fp16")];
fp16 _inversed_scores_17_y_0_to_fp16 = const()[name = string("_inversed_scores_17_y_0_to_fp16"), val = fp16(0x1p-3)];
tensor<fp16, [1, 8, 192, 192]> _inversed_scores_17_cast_fp16 = mul(x = var_801_cast_fp16, y = _inversed_scores_17_y_0_to_fp16)[name = string("_inversed_scores_17_cast_fp16")];
tensor<fp16, [1, 8, 192, 192]> scores_19_cast_fp16 = select(a = var_59_to_fp16, b = _inversed_scores_17_cast_fp16, cond = mask_3_cast_fp16)[name = string("scores_19_cast_fp16")];
tensor<fp16, [1, 8, 192, 192]> var_813_cast_fp16 = softmax(axis = var_77, x = scores_19_cast_fp16)[name = string("op_813_cast_fp16")];
tensor<fp16, [1, 8, 192, 192]> input_99_cast_fp16 = select(a = var_68_to_fp16, b = var_813_cast_fp16, cond = mask_3_cast_fp16)[name = string("input_99_cast_fp16")];
bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)];
bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)];
tensor<fp16, [1, 8, 192, 64]> v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = v_17_cast_fp16)[name = string("transpose_82")];
tensor<fp16, [1, 8, 192, 64]> x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = input_99_cast_fp16, y = v_19_cast_fp16)[name = string("x_13_cast_fp16")];
tensor<int32, [4]> var_817_perm_0 = const()[name = string("op_817_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<int32, [3]> var_819 = const()[name = string("op_819"), val = tensor<int32, [3]>([1, -1, 512])];
tensor<fp16, [1, 192, 8, 64]> var_817_cast_fp16 = transpose(perm = var_817_perm_0, x = x_13_cast_fp16)[name = string("transpose_78")];
tensor<fp16, [1, 192, 512]> input_101_cast_fp16 = reshape(shape = var_819, x = var_817_cast_fp16)[name = string("input_101_cast_fp16")];
tensor<fp16, [512, 512]> encoder_encoders_4_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_encoders_4_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39696320)))];
tensor<fp16, [512]> encoder_encoders_4_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_encoders_4_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40220672)))];
tensor<fp16, [1, 192, 512]> linear_33_cast_fp16 = linear(bias = encoder_encoders_4_self_attn_linear_out_bias_to_fp16, weight = encoder_encoders_4_self_attn_linear_out_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_33_cast_fp16")];
tensor<fp16, [1, 192, 512]> input_105_cast_fp16 = add(x = input_97_cast_fp16, y = linear_33_cast_fp16)[name = string("input_105_cast_fp16")];
tensor<int32, [1]> input_107_axes_0 = const()[name = string("input_107_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [512]> encoder_encoders_4_norm_ff_weight_to_fp16 = const()[name = string("encoder_encoders_4_norm_ff_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40221760)))];
tensor<fp16, [512]> encoder_encoders_4_norm_ff_bias_to_fp16 = const()[name = string("encoder_encoders_4_norm_ff_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40222848)))];
tensor<fp16, [1, 192, 512]> input_107_cast_fp16 = layer_norm(axes = input_107_axes_0, beta = encoder_encoders_4_norm_ff_bias_to_fp16, epsilon = var_58_to_fp16, gamma = encoder_encoders_4_norm_ff_weight_to_fp16, x = input_105_cast_fp16)[name = string("input_107_cast_fp16")];
tensor<fp16, [2048, 512]> encoder_encoders_4_feed_forward_w_1_weight_to_fp16 = const()[name = string("encoder_encoders_4_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40223936)))];
tensor<fp16, [2048]> encoder_encoders_4_feed_forward_w_1_bias_to_fp16 = const()[name = string("encoder_encoders_4_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42321152)))];
tensor<fp16, [1, 192, 2048]> linear_34_cast_fp16 = linear(bias = encoder_encoders_4_feed_forward_w_1_bias_to_fp16, weight = encoder_encoders_4_feed_forward_w_1_weight_to_fp16, x = input_107_cast_fp16)[name = string("linear_34_cast_fp16")];
tensor<fp16, [1, 192, 2048]> input_111_cast_fp16 = silu(x = linear_34_cast_fp16)[name = string("input_111_cast_fp16")];
tensor<fp16, [512, 2048]> encoder_encoders_4_feed_forward_w_2_weight_to_fp16 = const()[name = string("encoder_encoders_4_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42325312)))];
tensor<fp16, [512]> encoder_encoders_4_feed_forward_w_2_bias_to_fp16 = const()[name = string("encoder_encoders_4_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44422528)))];
tensor<fp16, [1, 192, 512]> linear_35_cast_fp16 = linear(bias = encoder_encoders_4_feed_forward_w_2_bias_to_fp16, weight = encoder_encoders_4_feed_forward_w_2_weight_to_fp16, x = input_111_cast_fp16)[name = string("linear_35_cast_fp16")];
tensor<fp16, [1, 192, 512]> input_117_cast_fp16 = add(x = input_105_cast_fp16, y = linear_35_cast_fp16)[name = string("input_117_cast_fp16")];
tensor<int32, [1]> query_11_axes_0 = const()[name = string("query_11_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [512]> encoder_encoders_5_norm_mha_weight_to_fp16 = const()[name = string("encoder_encoders_5_norm_mha_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44423616)))];
tensor<fp16, [512]> encoder_encoders_5_norm_mha_bias_to_fp16 = const()[name = string("encoder_encoders_5_norm_mha_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44424704)))];
tensor<fp16, [1, 192, 512]> query_11_cast_fp16 = layer_norm(axes = query_11_axes_0, beta = encoder_encoders_5_norm_mha_bias_to_fp16, epsilon = var_58_to_fp16, gamma = encoder_encoders_5_norm_mha_weight_to_fp16, x = input_117_cast_fp16)[name = string("query_11_cast_fp16")];
tensor<fp16, [512, 512]> encoder_encoders_5_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_encoders_5_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44425792)))];
tensor<fp16, [512]> encoder_encoders_5_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_encoders_5_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44950144)))];
tensor<fp16, [1, 192, 512]> linear_36_cast_fp16 = linear(bias = encoder_encoders_5_self_attn_linear_q_bias_to_fp16, weight = encoder_encoders_5_self_attn_linear_q_weight_to_fp16, x = query_11_cast_fp16)[name = string("linear_36_cast_fp16")];
tensor<int32, [4]> var_863 = const()[name = string("op_863"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 192, 8, 64]> q_31_cast_fp16 = reshape(shape = var_863, x = linear_36_cast_fp16)[name = string("q_31_cast_fp16")];
tensor<fp16, [512, 512]> encoder_encoders_5_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_encoders_5_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44951232)))];
tensor<fp16, [512]> encoder_encoders_5_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_encoders_5_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45475584)))];
tensor<fp16, [1, 192, 512]> linear_37_cast_fp16 = linear(bias = encoder_encoders_5_self_attn_linear_k_bias_to_fp16, weight = encoder_encoders_5_self_attn_linear_k_weight_to_fp16, x = query_11_cast_fp16)[name = string("linear_37_cast_fp16")];
tensor<int32, [4]> var_868 = const()[name = string("op_868"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 192, 8, 64]> k_21_cast_fp16 = reshape(shape = var_868, x = linear_37_cast_fp16)[name = string("k_21_cast_fp16")];
tensor<fp16, [512, 512]> encoder_encoders_5_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_encoders_5_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45476672)))];
tensor<fp16, [512]> encoder_encoders_5_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_encoders_5_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46001024)))];
tensor<fp16, [1, 192, 512]> linear_38_cast_fp16 = linear(bias = encoder_encoders_5_self_attn_linear_v_bias_to_fp16, weight = encoder_encoders_5_self_attn_linear_v_weight_to_fp16, x = query_11_cast_fp16)[name = string("linear_38_cast_fp16")];
tensor<int32, [4]> var_873 = const()[name = string("op_873"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 192, 8, 64]> v_21_cast_fp16 = reshape(shape = var_873, x = linear_38_cast_fp16)[name = string("v_21_cast_fp16")];
tensor<int32, [4]> v_23_perm_0 = const()[name = string("v_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<fp16, [8, 64]> const_103_to_fp16 = const()[name = string("const_103_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46002112)))];
tensor<fp16, [1, 192, 8, 64]> var_886_cast_fp16 = add(x = q_31_cast_fp16, y = const_103_to_fp16)[name = string("op_886_cast_fp16")];
tensor<fp16, [8, 64]> const_104_to_fp16 = const()[name = string("const_104_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46003200)))];
tensor<fp16, [1, 192, 8, 64]> var_889_cast_fp16 = add(x = q_31_cast_fp16, y = const_104_to_fp16)[name = string("op_889_cast_fp16")];
tensor<int32, [4]> q_with_bias_v_11_perm_0 = const()[name = string("q_with_bias_v_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
bool matrix_ac_11_transpose_x_0 = const()[name = string("matrix_ac_11_transpose_x_0"), val = bool(false)];
bool matrix_ac_11_transpose_y_0 = const()[name = string("matrix_ac_11_transpose_y_0"), val = bool(false)];
tensor<int32, [4]> transpose_40_perm_0 = const()[name = string("transpose_40_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
tensor<int32, [4]> transpose_41_perm_0 = const()[name = string("transpose_41_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
tensor<fp16, [1, 8, 64, 192]> transpose_41 = transpose(perm = transpose_41_perm_0, x = k_21_cast_fp16)[name = string("transpose_74")];
tensor<fp16, [1, 8, 192, 64]> transpose_40 = transpose(perm = transpose_40_perm_0, x = var_886_cast_fp16)[name = string("transpose_75")];
tensor<fp16, [1, 8, 192, 192]> matrix_ac_11_cast_fp16 = matmul(transpose_x = matrix_ac_11_transpose_x_0, transpose_y = matrix_ac_11_transpose_y_0, x = transpose_40, y = transpose_41)[name = string("matrix_ac_11_cast_fp16")];
bool matrix_bd_21_transpose_x_0 = const()[name = string("matrix_bd_21_transpose_x_0"), val = bool(false)];
bool matrix_bd_21_transpose_y_0 = const()[name = string("matrix_bd_21_transpose_y_0"), val = bool(false)];
tensor<fp16, [1, 8, 64, 383]> var_893_to_fp16 = const()[name = string("op_893_to_fp16"), val = tensor<fp16, [1, 8, 64, 383]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46004288)))];
tensor<fp16, [1, 8, 192, 64]> q_with_bias_v_11_cast_fp16 = transpose(perm = q_with_bias_v_11_perm_0, x = var_889_cast_fp16)[name = string("transpose_76")];
tensor<fp16, [1, 8, 192, 383]> matrix_bd_21_cast_fp16 = matmul(transpose_x = matrix_bd_21_transpose_x_0, transpose_y = matrix_bd_21_transpose_y_0, x = q_with_bias_v_11_cast_fp16, y = var_893_to_fp16)[name = string("matrix_bd_21_cast_fp16")];
bool x_padded_21_interleave_0 = const()[name = string("x_padded_21_interleave_0"), val = bool(false)];
tensor<fp16, [1, 8, 192, 384]> x_padded_21_cast_fp16 = concat(axis = var_77, interleave = x_padded_21_interleave_0, values = (zero_pad_1_to_fp16, matrix_bd_21_cast_fp16))[name = string("x_padded_21_cast_fp16")];
tensor<int32, [4]> var_910 = const()[name = string("op_910"), val = tensor<int32, [4]>([1, 8, 384, 192])];
tensor<fp16, [1, 8, 384, 192]> x_padded_23_cast_fp16 = reshape(shape = var_910, x = x_padded_21_cast_fp16)[name = string("x_padded_23_cast_fp16")];
tensor<int32, [4]> var_914_begin_0 = const()[name = string("op_914_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
tensor<int32, [4]> var_914_end_0 = const()[name = string("op_914_end_0"), val = tensor<int32, [4]>([1, 8, 384, 192])];
tensor<bool, [4]> var_914_end_mask_0 = const()[name = string("op_914_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
tensor<fp16, [1, 8, 383, 192]> var_914_cast_fp16 = slice_by_index(begin = var_914_begin_0, end = var_914_end_0, end_mask = var_914_end_mask_0, x = x_padded_23_cast_fp16)[name = string("op_914_cast_fp16")];
tensor<int32, [4]> var_919 = const()[name = string("op_919"), val = tensor<int32, [4]>([1, 8, 192, 383])];
tensor<fp16, [1, 8, 192, 383]> shifted_11_cast_fp16 = reshape(shape = var_919, x = var_914_cast_fp16)[name = string("shifted_11_cast_fp16")];
tensor<int32, [4]> matrix_bd_23_begin_0 = const()[name = string("matrix_bd_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [4]> matrix_bd_23_end_0 = const()[name = string("matrix_bd_23_end_0"), val = tensor<int32, [4]>([1, 8, 192, 192])];
tensor<bool, [4]> matrix_bd_23_end_mask_0 = const()[name = string("matrix_bd_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
tensor<fp16, [1, 8, 192, 192]> matrix_bd_23_cast_fp16 = slice_by_index(begin = matrix_bd_23_begin_0, end = matrix_bd_23_end_0, end_mask = matrix_bd_23_end_mask_0, x = shifted_11_cast_fp16)[name = string("matrix_bd_23_cast_fp16")];
tensor<fp16, [1, 8, 192, 192]> var_931_cast_fp16 = add(x = matrix_ac_11_cast_fp16, y = matrix_bd_23_cast_fp16)[name = string("op_931_cast_fp16")];
fp16 _inversed_scores_21_y_0_to_fp16 = const()[name = string("_inversed_scores_21_y_0_to_fp16"), val = fp16(0x1p-3)];
tensor<fp16, [1, 8, 192, 192]> _inversed_scores_21_cast_fp16 = mul(x = var_931_cast_fp16, y = _inversed_scores_21_y_0_to_fp16)[name = string("_inversed_scores_21_cast_fp16")];
tensor<fp16, [1, 8, 192, 192]> scores_23_cast_fp16 = select(a = var_59_to_fp16, b = _inversed_scores_21_cast_fp16, cond = mask_3_cast_fp16)[name = string("scores_23_cast_fp16")];
tensor<fp16, [1, 8, 192, 192]> var_943_cast_fp16 = softmax(axis = var_77, x = scores_23_cast_fp16)[name = string("op_943_cast_fp16")];
tensor<fp16, [1, 8, 192, 192]> input_119_cast_fp16 = select(a = var_68_to_fp16, b = var_943_cast_fp16, cond = mask_3_cast_fp16)[name = string("input_119_cast_fp16")];
bool x_15_transpose_x_0 = const()[name = string("x_15_transpose_x_0"), val = bool(false)];
bool x_15_transpose_y_0 = const()[name = string("x_15_transpose_y_0"), val = bool(false)];
tensor<fp16, [1, 8, 192, 64]> v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = v_21_cast_fp16)[name = string("transpose_77")];
tensor<fp16, [1, 8, 192, 64]> x_15_cast_fp16 = matmul(transpose_x = x_15_transpose_x_0, transpose_y = x_15_transpose_y_0, x = input_119_cast_fp16, y = v_23_cast_fp16)[name = string("x_15_cast_fp16")];
tensor<int32, [4]> var_947_perm_0 = const()[name = string("op_947_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<int32, [3]> var_949 = const()[name = string("op_949"), val = tensor<int32, [3]>([1, -1, 512])];
tensor<fp16, [1, 192, 8, 64]> var_947_cast_fp16 = transpose(perm = var_947_perm_0, x = x_15_cast_fp16)[name = string("transpose_73")];
tensor<fp16, [1, 192, 512]> input_121_cast_fp16 = reshape(shape = var_949, x = var_947_cast_fp16)[name = string("input_121_cast_fp16")];
tensor<fp16, [512, 512]> encoder_encoders_5_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_encoders_5_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46396544)))];
tensor<fp16, [512]> encoder_encoders_5_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_encoders_5_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46920896)))];
tensor<fp16, [1, 192, 512]> linear_40_cast_fp16 = linear(bias = encoder_encoders_5_self_attn_linear_out_bias_to_fp16, weight = encoder_encoders_5_self_attn_linear_out_weight_to_fp16, x = input_121_cast_fp16)[name = string("linear_40_cast_fp16")];
tensor<fp16, [1, 192, 512]> input_125_cast_fp16 = add(x = input_117_cast_fp16, y = linear_40_cast_fp16)[name = string("input_125_cast_fp16")];
tensor<int32, [1]> input_127_axes_0 = const()[name = string("input_127_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [512]> encoder_encoders_5_norm_ff_weight_to_fp16 = const()[name = string("encoder_encoders_5_norm_ff_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46921984)))];
tensor<fp16, [512]> encoder_encoders_5_norm_ff_bias_to_fp16 = const()[name = string("encoder_encoders_5_norm_ff_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46923072)))];
tensor<fp16, [1, 192, 512]> input_127_cast_fp16 = layer_norm(axes = input_127_axes_0, beta = encoder_encoders_5_norm_ff_bias_to_fp16, epsilon = var_58_to_fp16, gamma = encoder_encoders_5_norm_ff_weight_to_fp16, x = input_125_cast_fp16)[name = string("input_127_cast_fp16")];
tensor<fp16, [2048, 512]> encoder_encoders_5_feed_forward_w_1_weight_to_fp16 = const()[name = string("encoder_encoders_5_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46924160)))];
tensor<fp16, [2048]> encoder_encoders_5_feed_forward_w_1_bias_to_fp16 = const()[name = string("encoder_encoders_5_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49021376)))];
tensor<fp16, [1, 192, 2048]> linear_41_cast_fp16 = linear(bias = encoder_encoders_5_feed_forward_w_1_bias_to_fp16, weight = encoder_encoders_5_feed_forward_w_1_weight_to_fp16, x = input_127_cast_fp16)[name = string("linear_41_cast_fp16")];
tensor<fp16, [1, 192, 2048]> input_131_cast_fp16 = silu(x = linear_41_cast_fp16)[name = string("input_131_cast_fp16")];
tensor<fp16, [512, 2048]> encoder_encoders_5_feed_forward_w_2_weight_to_fp16 = const()[name = string("encoder_encoders_5_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49025536)))];
tensor<fp16, [512]> encoder_encoders_5_feed_forward_w_2_bias_to_fp16 = const()[name = string("encoder_encoders_5_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51122752)))];
tensor<fp16, [1, 192, 512]> linear_42_cast_fp16 = linear(bias = encoder_encoders_5_feed_forward_w_2_bias_to_fp16, weight = encoder_encoders_5_feed_forward_w_2_weight_to_fp16, x = input_131_cast_fp16)[name = string("linear_42_cast_fp16")];
tensor<fp16, [1, 192, 512]> xs_3_cast_fp16 = add(x = input_125_cast_fp16, y = linear_42_cast_fp16)[name = string("xs_3_cast_fp16")];
tensor<int32, [3]> var_974_perm_0 = const()[name = string("op_974_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
tensor<int32, [1]> expand_dims_0_axes_0 = const()[name = string("expand_dims_0_axes_0"), val = tensor<int32, [1]>([3])];
tensor<fp16, [1, 512, 192]> var_974_cast_fp16 = transpose(perm = var_974_perm_0, x = xs_3_cast_fp16)[name = string("transpose_72")];
tensor<fp16, [1, 512, 192, 1]> expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = var_974_cast_fp16)[name = string("expand_dims_0_cast_fp16")];
int32 upsample_nearest_neighbor_0_scale_factor_height_0 = const()[name = string("upsample_nearest_neighbor_0_scale_factor_height_0"), val = int32(2)];
int32 upsample_nearest_neighbor_0_scale_factor_width_0 = const()[name = string("upsample_nearest_neighbor_0_scale_factor_width_0"), val = int32(1)];
tensor<fp16, [1, 512, 384, 1]> upsample_nearest_neighbor_0_cast_fp16 = upsample_nearest_neighbor(scale_factor_height = upsample_nearest_neighbor_0_scale_factor_height_0, scale_factor_width = upsample_nearest_neighbor_0_scale_factor_width_0, x = expand_dims_0_cast_fp16)[name = string("upsample_nearest_neighbor_0_cast_fp16")];
tensor<int32, [1]> input_139_axes_0 = const()[name = string("input_139_axes_0"), val = tensor<int32, [1]>([3])];
tensor<fp16, [1, 512, 384]> input_139_cast_fp16 = squeeze(axes = input_139_axes_0, x = upsample_nearest_neighbor_0_cast_fp16)[name = string("input_139_cast_fp16")];
tensor<int32, [6]> input_141_pad_0 = const()[name = string("input_141_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 4, 0])];
string input_141_mode_0 = const()[name = string("input_141_mode_0"), val = string("constant")];
fp16 const_119_to_fp16 = const()[name = string("const_119_to_fp16"), val = fp16(0x0p+0)];
tensor<fp16, [1, 512, 388]> input_141_cast_fp16 = pad(constant_val = const_119_to_fp16, mode = input_141_mode_0, pad = input_141_pad_0, x = input_139_cast_fp16)[name = string("input_141_cast_fp16")];
string xs_5_pad_type_0 = const()[name = string("xs_5_pad_type_0"), val = string("valid")];
tensor<int32, [1]> xs_5_strides_0 = const()[name = string("xs_5_strides_0"), val = tensor<int32, [1]>([1])];
tensor<int32, [2]> xs_5_pad_0 = const()[name = string("xs_5_pad_0"), val = tensor<int32, [2]>([0, 0])];
tensor<int32, [1]> xs_5_dilations_0 = const()[name = string("xs_5_dilations_0"), val = tensor<int32, [1]>([1])];
int32 xs_5_groups_0 = const()[name = string("xs_5_groups_0"), val = int32(1)];
tensor<fp16, [512, 512, 5]> encoder_up_layer_conv_weight_to_fp16 = const()[name = string("encoder_up_layer_conv_weight_to_fp16"), val = tensor<fp16, [512, 512, 5]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51123840)))];
tensor<fp16, [512]> encoder_up_layer_conv_bias_to_fp16 = const()[name = string("encoder_up_layer_conv_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53745344)))];
tensor<fp16, [1, 512, 384]> xs_5_cast_fp16 = conv(bias = encoder_up_layer_conv_bias_to_fp16, dilations = xs_5_dilations_0, groups = xs_5_groups_0, pad = xs_5_pad_0, pad_type = xs_5_pad_type_0, strides = xs_5_strides_0, weight = encoder_up_layer_conv_weight_to_fp16, x = input_141_cast_fp16)[name = string("xs_5_cast_fp16")];
int32 var_988 = const()[name = string("op_988"), val = int32(2)];
tensor<int32, [1]> lengths_7 = mul(x = token_len, y = var_988)[name = string("lengths_7")];
tensor<int32, [3]> var_993_perm_0 = const()[name = string("op_993_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
tensor<int32, [1, 384]> var_999 = const()[name = string("op_999"), val = tensor<int32, [1, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53746432)))];
tensor<int32, [1]> seq_length_expand_axes_0 = const()[name = string("seq_length_expand_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<int32, [1, 1]> seq_length_expand = expand_dims(axes = seq_length_expand_axes_0, x = lengths_7)[name = string("seq_length_expand")];
tensor<bool, [1, 384]> var_1003 = greater_equal(x = var_999, y = seq_length_expand)[name = string("op_1003")];
tensor<int32, [1]> var_1004_axes_0 = const()[name = string("op_1004_axes_0"), val = tensor<int32, [1]>([1])];
tensor<bool, [1, 1, 384]> var_1004 = expand_dims(axes = var_1004_axes_0, x = var_1003)[name = string("op_1004")];
tensor<bool, [1, 1, 384]> masks = logical_not(x = var_1004)[name = string("masks")];
tensor<fp16, [512, 512]> encoder_up_embed_out_0_weight_to_fp16 = const()[name = string("encoder_up_embed_out_0_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53748032)))];
tensor<fp16, [512]> encoder_up_embed_out_0_bias_to_fp16 = const()[name = string("encoder_up_embed_out_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54272384)))];
tensor<fp16, [1, 384, 512]> var_993_cast_fp16 = transpose(perm = var_993_perm_0, x = xs_5_cast_fp16)[name = string("transpose_71")];
tensor<fp16, [1, 384, 512]> linear_43_cast_fp16 = linear(bias = encoder_up_embed_out_0_bias_to_fp16, weight = encoder_up_embed_out_0_weight_to_fp16, x = var_993_cast_fp16)[name = string("linear_43_cast_fp16")];
tensor<int32, [1]> input_145_axes_0 = const()[name = string("input_145_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [512]> encoder_up_embed_out_1_weight_to_fp16 = const()[name = string("encoder_up_embed_out_1_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54273472)))];
tensor<fp16, [512]> encoder_up_embed_out_1_bias_to_fp16 = const()[name = string("encoder_up_embed_out_1_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54274560)))];
tensor<fp16, [1, 384, 512]> input_145_cast_fp16 = layer_norm(axes = input_145_axes_0, beta = encoder_up_embed_out_1_bias_to_fp16, epsilon = var_72_to_fp16, gamma = encoder_up_embed_out_1_weight_to_fp16, x = linear_43_cast_fp16)[name = string("input_145_cast_fp16")];
fp16 var_1017_to_fp16 = const()[name = string("op_1017_to_fp16"), val = fp16(0x1.6ap+4)];
tensor<fp16, [1, 384, 512]> x_19_cast_fp16 = mul(x = input_145_cast_fp16, y = var_1017_to_fp16)[name = string("x_19_cast_fp16")];
tensor<int32, [1]> var_1044_axes_0 = const()[name = string("op_1044_axes_0"), val = tensor<int32, [1]>([-1])];
bool var_1044_keep_dims_0 = const()[name = string("op_1044_keep_dims_0"), val = bool(true)];
string cast_62_to_fp16_dtype_0 = const()[name = string("cast_62_to_fp16_dtype_0"), val = string("fp16")];
tensor<fp16, [1, 1, 384]> masks_to_fp16 = cast(dtype = cast_62_to_fp16_dtype_0, x = masks)[name = string("cast_103")];
tensor<fp16, [1, 1, 1]> var_1044_cast_fp16 = reduce_sum(axes = var_1044_axes_0, keep_dims = var_1044_keep_dims_0, x = masks_to_fp16)[name = string("op_1044_cast_fp16")];
fp16 var_79_promoted_7_to_fp16 = const()[name = string("op_79_promoted_7_to_fp16"), val = fp16(0x0p+0)];
tensor<bool, [1, 1, 1]> all_masked_cast_fp16 = equal(x = var_1044_cast_fp16, y = var_79_promoted_7_to_fp16)[name = string("all_masked_cast_fp16")];
fp16 fill_like_1_value_0_to_fp16 = const()[name = string("fill_like_1_value_0_to_fp16"), val = fp16(0x1p+0)];
tensor<fp16, [1, 1, 384]> fill_like_1_cast_fp16 = fill_like(ref_tensor = masks, value = fill_like_1_value_0_to_fp16)[name = string("fill_like_1_cast_fp16")];
tensor<int32, [3]> all_masked_after_broadcast_reps_0 = const()[name = string("all_masked_after_broadcast_reps_0"), val = tensor<int32, [3]>([1, 1, 384])];
tensor<bool, [1, 1, 384]> all_masked_after_broadcast = tile(reps = all_masked_after_broadcast_reps_0, x = all_masked_cast_fp16)[name = string("all_masked_after_broadcast")];
tensor<fp16, [1, 1, 384]> mask_27_cast_fp16 = select(a = fill_like_1_cast_fp16, b = masks_to_fp16, cond = all_masked_after_broadcast)[name = string("mask_27_cast_fp16")];
tensor<int32, [1]> query_13_axes_0 = const()[name = string("query_13_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [512]> encoder_up_encoders_0_norm_mha_weight_to_fp16 = const()[name = string("encoder_up_encoders_0_norm_mha_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54275648)))];
tensor<fp16, [512]> encoder_up_encoders_0_norm_mha_bias_to_fp16 = const()[name = string("encoder_up_encoders_0_norm_mha_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54276736)))];
tensor<fp16, [1, 384, 512]> query_13_cast_fp16 = layer_norm(axes = query_13_axes_0, beta = encoder_up_encoders_0_norm_mha_bias_to_fp16, epsilon = var_58_to_fp16, gamma = encoder_up_encoders_0_norm_mha_weight_to_fp16, x = x_19_cast_fp16)[name = string("query_13_cast_fp16")];
tensor<fp16, [512, 512]> encoder_up_encoders_0_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_up_encoders_0_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54277824)))];
tensor<fp16, [512]> encoder_up_encoders_0_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_up_encoders_0_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54802176)))];
tensor<fp16, [1, 384, 512]> linear_44_cast_fp16 = linear(bias = encoder_up_encoders_0_self_attn_linear_q_bias_to_fp16, weight = encoder_up_encoders_0_self_attn_linear_q_weight_to_fp16, x = query_13_cast_fp16)[name = string("linear_44_cast_fp16")];
tensor<int32, [4]> var_1067 = const()[name = string("op_1067"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 384, 8, 64]> q_37_cast_fp16 = reshape(shape = var_1067, x = linear_44_cast_fp16)[name = string("q_37_cast_fp16")];
tensor<fp16, [512, 512]> encoder_up_encoders_0_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_up_encoders_0_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54803264)))];
tensor<fp16, [512]> encoder_up_encoders_0_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_up_encoders_0_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55327616)))];
tensor<fp16, [1, 384, 512]> linear_45_cast_fp16 = linear(bias = encoder_up_encoders_0_self_attn_linear_k_bias_to_fp16, weight = encoder_up_encoders_0_self_attn_linear_k_weight_to_fp16, x = query_13_cast_fp16)[name = string("linear_45_cast_fp16")];
tensor<int32, [4]> var_1072 = const()[name = string("op_1072"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 384, 8, 64]> k_25_cast_fp16 = reshape(shape = var_1072, x = linear_45_cast_fp16)[name = string("k_25_cast_fp16")];
tensor<fp16, [512, 512]> encoder_up_encoders_0_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_up_encoders_0_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55328704)))];
tensor<fp16, [512]> encoder_up_encoders_0_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_up_encoders_0_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55853056)))];
tensor<fp16, [1, 384, 512]> linear_46_cast_fp16 = linear(bias = encoder_up_encoders_0_self_attn_linear_v_bias_to_fp16, weight = encoder_up_encoders_0_self_attn_linear_v_weight_to_fp16, x = query_13_cast_fp16)[name = string("linear_46_cast_fp16")];
tensor<int32, [4]> var_1077 = const()[name = string("op_1077"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 384, 8, 64]> v_25_cast_fp16 = reshape(shape = var_1077, x = linear_46_cast_fp16)[name = string("v_25_cast_fp16")];
tensor<int32, [4]> v_27_perm_0 = const()[name = string("v_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<fp16, [8, 64]> const_128_to_fp16 = const()[name = string("const_128_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55854144)))];
tensor<fp16, [1, 384, 8, 64]> var_1090_cast_fp16 = add(x = q_37_cast_fp16, y = const_128_to_fp16)[name = string("op_1090_cast_fp16")];
tensor<fp16, [8, 64]> const_129_to_fp16 = const()[name = string("const_129_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55855232)))];
tensor<fp16, [1, 384, 8, 64]> var_1093_cast_fp16 = add(x = q_37_cast_fp16, y = const_129_to_fp16)[name = string("op_1093_cast_fp16")];
tensor<int32, [4]> q_with_bias_v_13_perm_0 = const()[name = string("q_with_bias_v_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
bool matrix_ac_13_transpose_x_0 = const()[name = string("matrix_ac_13_transpose_x_0"), val = bool(false)];
bool matrix_ac_13_transpose_y_0 = const()[name = string("matrix_ac_13_transpose_y_0"), val = bool(false)];
tensor<int32, [4]> transpose_42_perm_0 = const()[name = string("transpose_42_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
tensor<int32, [4]> transpose_43_perm_0 = const()[name = string("transpose_43_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
tensor<fp16, [1, 8, 64, 384]> transpose_43 = transpose(perm = transpose_43_perm_0, x = k_25_cast_fp16)[name = string("transpose_67")];
tensor<fp16, [1, 8, 384, 64]> transpose_42 = transpose(perm = transpose_42_perm_0, x = var_1090_cast_fp16)[name = string("transpose_68")];
tensor<fp16, [1, 8, 384, 384]> matrix_ac_13_cast_fp16 = matmul(transpose_x = matrix_ac_13_transpose_x_0, transpose_y = matrix_ac_13_transpose_y_0, x = transpose_42, y = transpose_43)[name = string("matrix_ac_13_cast_fp16")];
bool matrix_bd_25_transpose_x_0 = const()[name = string("matrix_bd_25_transpose_x_0"), val = bool(false)];
bool matrix_bd_25_transpose_y_0 = const()[name = string("matrix_bd_25_transpose_y_0"), val = bool(false)];
tensor<fp16, [1, 8, 64, 767]> var_1097_to_fp16 = const()[name = string("op_1097_to_fp16"), val = tensor<fp16, [1, 8, 64, 767]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55856320)))];
tensor<fp16, [1, 8, 384, 64]> q_with_bias_v_13_cast_fp16 = transpose(perm = q_with_bias_v_13_perm_0, x = var_1093_cast_fp16)[name = string("transpose_69")];
tensor<fp16, [1, 8, 384, 767]> matrix_bd_25_cast_fp16 = matmul(transpose_x = matrix_bd_25_transpose_x_0, transpose_y = matrix_bd_25_transpose_y_0, x = q_with_bias_v_13_cast_fp16, y = var_1097_to_fp16)[name = string("matrix_bd_25_cast_fp16")];
bool x_padded_25_interleave_0 = const()[name = string("x_padded_25_interleave_0"), val = bool(false)];
tensor<fp16, [1, 8, 384, 1]> zero_pad_13_to_fp16 = const()[name = string("zero_pad_13_to_fp16"), val = tensor<fp16, [1, 8, 384, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56641792)))];
tensor<fp16, [1, 8, 384, 768]> x_padded_25_cast_fp16 = concat(axis = var_77, interleave = x_padded_25_interleave_0, values = (zero_pad_13_to_fp16, matrix_bd_25_cast_fp16))[name = string("x_padded_25_cast_fp16")];
tensor<int32, [4]> var_1114 = const()[name = string("op_1114"), val = tensor<int32, [4]>([1, 8, 768, 384])];
tensor<fp16, [1, 8, 768, 384]> x_padded_27_cast_fp16 = reshape(shape = var_1114, x = x_padded_25_cast_fp16)[name = string("x_padded_27_cast_fp16")];
tensor<int32, [4]> var_1118_begin_0 = const()[name = string("op_1118_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
tensor<int32, [4]> var_1118_end_0 = const()[name = string("op_1118_end_0"), val = tensor<int32, [4]>([1, 8, 768, 384])];
tensor<bool, [4]> var_1118_end_mask_0 = const()[name = string("op_1118_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
tensor<fp16, [1, 8, 767, 384]> var_1118_cast_fp16 = slice_by_index(begin = var_1118_begin_0, end = var_1118_end_0, end_mask = var_1118_end_mask_0, x = x_padded_27_cast_fp16)[name = string("op_1118_cast_fp16")];
tensor<int32, [4]> var_1123 = const()[name = string("op_1123"), val = tensor<int32, [4]>([1, 8, 384, 767])];
tensor<fp16, [1, 8, 384, 767]> shifted_13_cast_fp16 = reshape(shape = var_1123, x = var_1118_cast_fp16)[name = string("shifted_13_cast_fp16")];
tensor<int32, [4]> matrix_bd_27_begin_0 = const()[name = string("matrix_bd_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [4]> matrix_bd_27_end_0 = const()[name = string("matrix_bd_27_end_0"), val = tensor<int32, [4]>([1, 8, 384, 384])];
tensor<bool, [4]> matrix_bd_27_end_mask_0 = const()[name = string("matrix_bd_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
tensor<fp16, [1, 8, 384, 384]> matrix_bd_27_cast_fp16 = slice_by_index(begin = matrix_bd_27_begin_0, end = matrix_bd_27_end_0, end_mask = matrix_bd_27_end_mask_0, x = shifted_13_cast_fp16)[name = string("matrix_bd_27_cast_fp16")];
tensor<fp16, [1, 8, 384, 384]> var_1135_cast_fp16 = add(x = matrix_ac_13_cast_fp16, y = matrix_bd_27_cast_fp16)[name = string("op_1135_cast_fp16")];
fp16 _inversed_scores_25_y_0_to_fp16 = const()[name = string("_inversed_scores_25_y_0_to_fp16"), val = fp16(0x1p-3)];
tensor<fp16, [1, 8, 384, 384]> _inversed_scores_25_cast_fp16 = mul(x = var_1135_cast_fp16, y = _inversed_scores_25_y_0_to_fp16)[name = string("_inversed_scores_25_cast_fp16")];
tensor<int32, [1]> var_1139_axes_0 = const()[name = string("op_1139_axes_0"), val = tensor<int32, [1]>([1])];
tensor<fp16, [1, 1, 1, 384]> var_1139_cast_fp16 = expand_dims(axes = var_1139_axes_0, x = mask_27_cast_fp16)[name = string("op_1139_cast_fp16")];
fp16 var_79_promoted_8_to_fp16 = const()[name = string("op_79_promoted_8_to_fp16"), val = fp16(0x0p+0)];
tensor<bool, [1, 1, 1, 384]> mask_29_cast_fp16 = equal(x = var_1139_cast_fp16, y = var_79_promoted_8_to_fp16)[name = string("mask_29_cast_fp16")];
tensor<fp16, [1, 8, 384, 384]> scores_27_cast_fp16 = select(a = var_59_to_fp16, b = _inversed_scores_25_cast_fp16, cond = mask_29_cast_fp16)[name = string("scores_27_cast_fp16")];
tensor<fp16, [1, 8, 384, 384]> var_1147_cast_fp16 = softmax(axis = var_77, x = scores_27_cast_fp16)[name = string("op_1147_cast_fp16")];
tensor<fp16, [1, 8, 384, 384]> input_151_cast_fp16 = select(a = var_68_to_fp16, b = var_1147_cast_fp16, cond = mask_29_cast_fp16)[name = string("input_151_cast_fp16")];
bool x_21_transpose_x_0 = const()[name = string("x_21_transpose_x_0"), val = bool(false)];
bool x_21_transpose_y_0 = const()[name = string("x_21_transpose_y_0"), val = bool(false)];
tensor<fp16, [1, 8, 384, 64]> v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = v_25_cast_fp16)[name = string("transpose_70")];
tensor<fp16, [1, 8, 384, 64]> x_21_cast_fp16 = matmul(transpose_x = x_21_transpose_x_0, transpose_y = x_21_transpose_y_0, x = input_151_cast_fp16, y = v_27_cast_fp16)[name = string("x_21_cast_fp16")];
tensor<int32, [4]> var_1151_perm_0 = const()[name = string("op_1151_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<int32, [3]> var_1153 = const()[name = string("op_1153"), val = tensor<int32, [3]>([1, -1, 512])];
tensor<fp16, [1, 384, 8, 64]> var_1151_cast_fp16 = transpose(perm = var_1151_perm_0, x = x_21_cast_fp16)[name = string("transpose_66")];
tensor<fp16, [1, 384, 512]> input_153_cast_fp16 = reshape(shape = var_1153, x = var_1151_cast_fp16)[name = string("input_153_cast_fp16")];
tensor<fp16, [512, 512]> encoder_up_encoders_0_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_up_encoders_0_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56648000)))];
tensor<fp16, [512]> encoder_up_encoders_0_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_up_encoders_0_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57172352)))];
tensor<fp16, [1, 384, 512]> linear_48_cast_fp16 = linear(bias = encoder_up_encoders_0_self_attn_linear_out_bias_to_fp16, weight = encoder_up_encoders_0_self_attn_linear_out_weight_to_fp16, x = input_153_cast_fp16)[name = string("linear_48_cast_fp16")];
tensor<fp16, [1, 384, 512]> input_157_cast_fp16 = add(x = x_19_cast_fp16, y = linear_48_cast_fp16)[name = string("input_157_cast_fp16")];
tensor<int32, [1]> input_159_axes_0 = const()[name = string("input_159_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [512]> encoder_up_encoders_0_norm_ff_weight_to_fp16 = const()[name = string("encoder_up_encoders_0_norm_ff_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57173440)))];
tensor<fp16, [512]> encoder_up_encoders_0_norm_ff_bias_to_fp16 = const()[name = string("encoder_up_encoders_0_norm_ff_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57174528)))];
tensor<fp16, [1, 384, 512]> input_159_cast_fp16 = layer_norm(axes = input_159_axes_0, beta = encoder_up_encoders_0_norm_ff_bias_to_fp16, epsilon = var_58_to_fp16, gamma = encoder_up_encoders_0_norm_ff_weight_to_fp16, x = input_157_cast_fp16)[name = string("input_159_cast_fp16")];
tensor<fp16, [2048, 512]> encoder_up_encoders_0_feed_forward_w_1_weight_to_fp16 = const()[name = string("encoder_up_encoders_0_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57175616)))];
tensor<fp16, [2048]> encoder_up_encoders_0_feed_forward_w_1_bias_to_fp16 = const()[name = string("encoder_up_encoders_0_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59272832)))];
tensor<fp16, [1, 384, 2048]> linear_49_cast_fp16 = linear(bias = encoder_up_encoders_0_feed_forward_w_1_bias_to_fp16, weight = encoder_up_encoders_0_feed_forward_w_1_weight_to_fp16, x = input_159_cast_fp16)[name = string("linear_49_cast_fp16")];
tensor<fp16, [1, 384, 2048]> input_163_cast_fp16 = silu(x = linear_49_cast_fp16)[name = string("input_163_cast_fp16")];
tensor<fp16, [512, 2048]> encoder_up_encoders_0_feed_forward_w_2_weight_to_fp16 = const()[name = string("encoder_up_encoders_0_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59276992)))];
tensor<fp16, [512]> encoder_up_encoders_0_feed_forward_w_2_bias_to_fp16 = const()[name = string("encoder_up_encoders_0_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61374208)))];
tensor<fp16, [1, 384, 512]> linear_50_cast_fp16 = linear(bias = encoder_up_encoders_0_feed_forward_w_2_bias_to_fp16, weight = encoder_up_encoders_0_feed_forward_w_2_weight_to_fp16, x = input_163_cast_fp16)[name = string("linear_50_cast_fp16")];
tensor<fp16, [1, 384, 512]> input_169_cast_fp16 = add(x = input_157_cast_fp16, y = linear_50_cast_fp16)[name = string("input_169_cast_fp16")];
tensor<int32, [1]> query_15_axes_0 = const()[name = string("query_15_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [512]> encoder_up_encoders_1_norm_mha_weight_to_fp16 = const()[name = string("encoder_up_encoders_1_norm_mha_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61375296)))];
tensor<fp16, [512]> encoder_up_encoders_1_norm_mha_bias_to_fp16 = const()[name = string("encoder_up_encoders_1_norm_mha_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61376384)))];
tensor<fp16, [1, 384, 512]> query_15_cast_fp16 = layer_norm(axes = query_15_axes_0, beta = encoder_up_encoders_1_norm_mha_bias_to_fp16, epsilon = var_58_to_fp16, gamma = encoder_up_encoders_1_norm_mha_weight_to_fp16, x = input_169_cast_fp16)[name = string("query_15_cast_fp16")];
tensor<fp16, [512, 512]> encoder_up_encoders_1_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_up_encoders_1_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61377472)))];
tensor<fp16, [512]> encoder_up_encoders_1_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_up_encoders_1_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61901824)))];
tensor<fp16, [1, 384, 512]> linear_51_cast_fp16 = linear(bias = encoder_up_encoders_1_self_attn_linear_q_bias_to_fp16, weight = encoder_up_encoders_1_self_attn_linear_q_weight_to_fp16, x = query_15_cast_fp16)[name = string("linear_51_cast_fp16")];
tensor<int32, [4]> var_1197 = const()[name = string("op_1197"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 384, 8, 64]> q_43_cast_fp16 = reshape(shape = var_1197, x = linear_51_cast_fp16)[name = string("q_43_cast_fp16")];
tensor<fp16, [512, 512]> encoder_up_encoders_1_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_up_encoders_1_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61902912)))];
tensor<fp16, [512]> encoder_up_encoders_1_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_up_encoders_1_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62427264)))];
tensor<fp16, [1, 384, 512]> linear_52_cast_fp16 = linear(bias = encoder_up_encoders_1_self_attn_linear_k_bias_to_fp16, weight = encoder_up_encoders_1_self_attn_linear_k_weight_to_fp16, x = query_15_cast_fp16)[name = string("linear_52_cast_fp16")];
tensor<int32, [4]> var_1202 = const()[name = string("op_1202"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 384, 8, 64]> k_29_cast_fp16 = reshape(shape = var_1202, x = linear_52_cast_fp16)[name = string("k_29_cast_fp16")];
tensor<fp16, [512, 512]> encoder_up_encoders_1_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_up_encoders_1_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62428352)))];
tensor<fp16, [512]> encoder_up_encoders_1_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_up_encoders_1_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62952704)))];
tensor<fp16, [1, 384, 512]> linear_53_cast_fp16 = linear(bias = encoder_up_encoders_1_self_attn_linear_v_bias_to_fp16, weight = encoder_up_encoders_1_self_attn_linear_v_weight_to_fp16, x = query_15_cast_fp16)[name = string("linear_53_cast_fp16")];
tensor<int32, [4]> var_1207 = const()[name = string("op_1207"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 384, 8, 64]> v_29_cast_fp16 = reshape(shape = var_1207, x = linear_53_cast_fp16)[name = string("v_29_cast_fp16")];
tensor<int32, [4]> v_31_perm_0 = const()[name = string("v_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<fp16, [8, 64]> const_146_to_fp16 = const()[name = string("const_146_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62953792)))];
tensor<fp16, [1, 384, 8, 64]> var_1220_cast_fp16 = add(x = q_43_cast_fp16, y = const_146_to_fp16)[name = string("op_1220_cast_fp16")];
tensor<fp16, [8, 64]> const_147_to_fp16 = const()[name = string("const_147_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62954880)))];
tensor<fp16, [1, 384, 8, 64]> var_1223_cast_fp16 = add(x = q_43_cast_fp16, y = const_147_to_fp16)[name = string("op_1223_cast_fp16")];
tensor<int32, [4]> q_with_bias_v_15_perm_0 = const()[name = string("q_with_bias_v_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
bool matrix_ac_15_transpose_x_0 = const()[name = string("matrix_ac_15_transpose_x_0"), val = bool(false)];
bool matrix_ac_15_transpose_y_0 = const()[name = string("matrix_ac_15_transpose_y_0"), val = bool(false)];
tensor<int32, [4]> transpose_44_perm_0 = const()[name = string("transpose_44_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
tensor<int32, [4]> transpose_45_perm_0 = const()[name = string("transpose_45_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
tensor<fp16, [1, 8, 64, 384]> transpose_45 = transpose(perm = transpose_45_perm_0, x = k_29_cast_fp16)[name = string("transpose_62")];
tensor<fp16, [1, 8, 384, 64]> transpose_44 = transpose(perm = transpose_44_perm_0, x = var_1220_cast_fp16)[name = string("transpose_63")];
tensor<fp16, [1, 8, 384, 384]> matrix_ac_15_cast_fp16 = matmul(transpose_x = matrix_ac_15_transpose_x_0, transpose_y = matrix_ac_15_transpose_y_0, x = transpose_44, y = transpose_45)[name = string("matrix_ac_15_cast_fp16")];
bool matrix_bd_29_transpose_x_0 = const()[name = string("matrix_bd_29_transpose_x_0"), val = bool(false)];
bool matrix_bd_29_transpose_y_0 = const()[name = string("matrix_bd_29_transpose_y_0"), val = bool(false)];
tensor<fp16, [1, 8, 64, 767]> var_1227_to_fp16 = const()[name = string("op_1227_to_fp16"), val = tensor<fp16, [1, 8, 64, 767]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62955968)))];
tensor<fp16, [1, 8, 384, 64]> q_with_bias_v_15_cast_fp16 = transpose(perm = q_with_bias_v_15_perm_0, x = var_1223_cast_fp16)[name = string("transpose_64")];
tensor<fp16, [1, 8, 384, 767]> matrix_bd_29_cast_fp16 = matmul(transpose_x = matrix_bd_29_transpose_x_0, transpose_y = matrix_bd_29_transpose_y_0, x = q_with_bias_v_15_cast_fp16, y = var_1227_to_fp16)[name = string("matrix_bd_29_cast_fp16")];
bool x_padded_29_interleave_0 = const()[name = string("x_padded_29_interleave_0"), val = bool(false)];
tensor<fp16, [1, 8, 384, 768]> x_padded_29_cast_fp16 = concat(axis = var_77, interleave = x_padded_29_interleave_0, values = (zero_pad_13_to_fp16, matrix_bd_29_cast_fp16))[name = string("x_padded_29_cast_fp16")];
tensor<int32, [4]> var_1244 = const()[name = string("op_1244"), val = tensor<int32, [4]>([1, 8, 768, 384])];
tensor<fp16, [1, 8, 768, 384]> x_padded_31_cast_fp16 = reshape(shape = var_1244, x = x_padded_29_cast_fp16)[name = string("x_padded_31_cast_fp16")];
tensor<int32, [4]> var_1248_begin_0 = const()[name = string("op_1248_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
tensor<int32, [4]> var_1248_end_0 = const()[name = string("op_1248_end_0"), val = tensor<int32, [4]>([1, 8, 768, 384])];
tensor<bool, [4]> var_1248_end_mask_0 = const()[name = string("op_1248_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
tensor<fp16, [1, 8, 767, 384]> var_1248_cast_fp16 = slice_by_index(begin = var_1248_begin_0, end = var_1248_end_0, end_mask = var_1248_end_mask_0, x = x_padded_31_cast_fp16)[name = string("op_1248_cast_fp16")];
tensor<int32, [4]> var_1253 = const()[name = string("op_1253"), val = tensor<int32, [4]>([1, 8, 384, 767])];
tensor<fp16, [1, 8, 384, 767]> shifted_15_cast_fp16 = reshape(shape = var_1253, x = var_1248_cast_fp16)[name = string("shifted_15_cast_fp16")];
tensor<int32, [4]> matrix_bd_31_begin_0 = const()[name = string("matrix_bd_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [4]> matrix_bd_31_end_0 = const()[name = string("matrix_bd_31_end_0"), val = tensor<int32, [4]>([1, 8, 384, 384])];
tensor<bool, [4]> matrix_bd_31_end_mask_0 = const()[name = string("matrix_bd_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
tensor<fp16, [1, 8, 384, 384]> matrix_bd_31_cast_fp16 = slice_by_index(begin = matrix_bd_31_begin_0, end = matrix_bd_31_end_0, end_mask = matrix_bd_31_end_mask_0, x = shifted_15_cast_fp16)[name = string("matrix_bd_31_cast_fp16")];
tensor<fp16, [1, 8, 384, 384]> var_1265_cast_fp16 = add(x = matrix_ac_15_cast_fp16, y = matrix_bd_31_cast_fp16)[name = string("op_1265_cast_fp16")];
fp16 _inversed_scores_29_y_0_to_fp16 = const()[name = string("_inversed_scores_29_y_0_to_fp16"), val = fp16(0x1p-3)];
tensor<fp16, [1, 8, 384, 384]> _inversed_scores_29_cast_fp16 = mul(x = var_1265_cast_fp16, y = _inversed_scores_29_y_0_to_fp16)[name = string("_inversed_scores_29_cast_fp16")];
tensor<fp16, [1, 8, 384, 384]> scores_31_cast_fp16 = select(a = var_59_to_fp16, b = _inversed_scores_29_cast_fp16, cond = mask_29_cast_fp16)[name = string("scores_31_cast_fp16")];
tensor<fp16, [1, 8, 384, 384]> var_1277_cast_fp16 = softmax(axis = var_77, x = scores_31_cast_fp16)[name = string("op_1277_cast_fp16")];
tensor<fp16, [1, 8, 384, 384]> input_171_cast_fp16 = select(a = var_68_to_fp16, b = var_1277_cast_fp16, cond = mask_29_cast_fp16)[name = string("input_171_cast_fp16")];
bool x_23_transpose_x_0 = const()[name = string("x_23_transpose_x_0"), val = bool(false)];
bool x_23_transpose_y_0 = const()[name = string("x_23_transpose_y_0"), val = bool(false)];
tensor<fp16, [1, 8, 384, 64]> v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = v_29_cast_fp16)[name = string("transpose_65")];
tensor<fp16, [1, 8, 384, 64]> x_23_cast_fp16 = matmul(transpose_x = x_23_transpose_x_0, transpose_y = x_23_transpose_y_0, x = input_171_cast_fp16, y = v_31_cast_fp16)[name = string("x_23_cast_fp16")];
tensor<int32, [4]> var_1281_perm_0 = const()[name = string("op_1281_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<int32, [3]> var_1283 = const()[name = string("op_1283"), val = tensor<int32, [3]>([1, -1, 512])];
tensor<fp16, [1, 384, 8, 64]> var_1281_cast_fp16 = transpose(perm = var_1281_perm_0, x = x_23_cast_fp16)[name = string("transpose_61")];
tensor<fp16, [1, 384, 512]> input_173_cast_fp16 = reshape(shape = var_1283, x = var_1281_cast_fp16)[name = string("input_173_cast_fp16")];
tensor<fp16, [512, 512]> encoder_up_encoders_1_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_up_encoders_1_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63741440)))];
tensor<fp16, [512]> encoder_up_encoders_1_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_up_encoders_1_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64265792)))];
tensor<fp16, [1, 384, 512]> linear_55_cast_fp16 = linear(bias = encoder_up_encoders_1_self_attn_linear_out_bias_to_fp16, weight = encoder_up_encoders_1_self_attn_linear_out_weight_to_fp16, x = input_173_cast_fp16)[name = string("linear_55_cast_fp16")];
tensor<fp16, [1, 384, 512]> input_177_cast_fp16 = add(x = input_169_cast_fp16, y = linear_55_cast_fp16)[name = string("input_177_cast_fp16")];
tensor<int32, [1]> input_179_axes_0 = const()[name = string("input_179_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [512]> encoder_up_encoders_1_norm_ff_weight_to_fp16 = const()[name = string("encoder_up_encoders_1_norm_ff_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64266880)))];
tensor<fp16, [512]> encoder_up_encoders_1_norm_ff_bias_to_fp16 = const()[name = string("encoder_up_encoders_1_norm_ff_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64267968)))];
tensor<fp16, [1, 384, 512]> input_179_cast_fp16 = layer_norm(axes = input_179_axes_0, beta = encoder_up_encoders_1_norm_ff_bias_to_fp16, epsilon = var_58_to_fp16, gamma = encoder_up_encoders_1_norm_ff_weight_to_fp16, x = input_177_cast_fp16)[name = string("input_179_cast_fp16")];
tensor<fp16, [2048, 512]> encoder_up_encoders_1_feed_forward_w_1_weight_to_fp16 = const()[name = string("encoder_up_encoders_1_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64269056)))];
tensor<fp16, [2048]> encoder_up_encoders_1_feed_forward_w_1_bias_to_fp16 = const()[name = string("encoder_up_encoders_1_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66366272)))];
tensor<fp16, [1, 384, 2048]> linear_56_cast_fp16 = linear(bias = encoder_up_encoders_1_feed_forward_w_1_bias_to_fp16, weight = encoder_up_encoders_1_feed_forward_w_1_weight_to_fp16, x = input_179_cast_fp16)[name = string("linear_56_cast_fp16")];
tensor<fp16, [1, 384, 2048]> input_183_cast_fp16 = silu(x = linear_56_cast_fp16)[name = string("input_183_cast_fp16")];
tensor<fp16, [512, 2048]> encoder_up_encoders_1_feed_forward_w_2_weight_to_fp16 = const()[name = string("encoder_up_encoders_1_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66370432)))];
tensor<fp16, [512]> encoder_up_encoders_1_feed_forward_w_2_bias_to_fp16 = const()[name = string("encoder_up_encoders_1_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68467648)))];
tensor<fp16, [1, 384, 512]> linear_57_cast_fp16 = linear(bias = encoder_up_encoders_1_feed_forward_w_2_bias_to_fp16, weight = encoder_up_encoders_1_feed_forward_w_2_weight_to_fp16, x = input_183_cast_fp16)[name = string("linear_57_cast_fp16")];
tensor<fp16, [1, 384, 512]> input_189_cast_fp16 = add(x = input_177_cast_fp16, y = linear_57_cast_fp16)[name = string("input_189_cast_fp16")];
tensor<int32, [1]> query_17_axes_0 = const()[name = string("query_17_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [512]> encoder_up_encoders_2_norm_mha_weight_to_fp16 = const()[name = string("encoder_up_encoders_2_norm_mha_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68468736)))];
tensor<fp16, [512]> encoder_up_encoders_2_norm_mha_bias_to_fp16 = const()[name = string("encoder_up_encoders_2_norm_mha_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68469824)))];
tensor<fp16, [1, 384, 512]> query_17_cast_fp16 = layer_norm(axes = query_17_axes_0, beta = encoder_up_encoders_2_norm_mha_bias_to_fp16, epsilon = var_58_to_fp16, gamma = encoder_up_encoders_2_norm_mha_weight_to_fp16, x = input_189_cast_fp16)[name = string("query_17_cast_fp16")];
tensor<fp16, [512, 512]> encoder_up_encoders_2_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_up_encoders_2_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68470912)))];
tensor<fp16, [512]> encoder_up_encoders_2_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_up_encoders_2_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68995264)))];
tensor<fp16, [1, 384, 512]> linear_58_cast_fp16 = linear(bias = encoder_up_encoders_2_self_attn_linear_q_bias_to_fp16, weight = encoder_up_encoders_2_self_attn_linear_q_weight_to_fp16, x = query_17_cast_fp16)[name = string("linear_58_cast_fp16")];
tensor<int32, [4]> var_1327 = const()[name = string("op_1327"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 384, 8, 64]> q_49_cast_fp16 = reshape(shape = var_1327, x = linear_58_cast_fp16)[name = string("q_49_cast_fp16")];
tensor<fp16, [512, 512]> encoder_up_encoders_2_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_up_encoders_2_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68996352)))];
tensor<fp16, [512]> encoder_up_encoders_2_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_up_encoders_2_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69520704)))];
tensor<fp16, [1, 384, 512]> linear_59_cast_fp16 = linear(bias = encoder_up_encoders_2_self_attn_linear_k_bias_to_fp16, weight = encoder_up_encoders_2_self_attn_linear_k_weight_to_fp16, x = query_17_cast_fp16)[name = string("linear_59_cast_fp16")];
tensor<int32, [4]> var_1332 = const()[name = string("op_1332"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 384, 8, 64]> k_33_cast_fp16 = reshape(shape = var_1332, x = linear_59_cast_fp16)[name = string("k_33_cast_fp16")];
tensor<fp16, [512, 512]> encoder_up_encoders_2_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_up_encoders_2_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69521792)))];
tensor<fp16, [512]> encoder_up_encoders_2_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_up_encoders_2_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70046144)))];
tensor<fp16, [1, 384, 512]> linear_60_cast_fp16 = linear(bias = encoder_up_encoders_2_self_attn_linear_v_bias_to_fp16, weight = encoder_up_encoders_2_self_attn_linear_v_weight_to_fp16, x = query_17_cast_fp16)[name = string("linear_60_cast_fp16")];
tensor<int32, [4]> var_1337 = const()[name = string("op_1337"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 384, 8, 64]> v_33_cast_fp16 = reshape(shape = var_1337, x = linear_60_cast_fp16)[name = string("v_33_cast_fp16")];
tensor<int32, [4]> v_35_perm_0 = const()[name = string("v_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<fp16, [8, 64]> const_164_to_fp16 = const()[name = string("const_164_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70047232)))];
tensor<fp16, [1, 384, 8, 64]> var_1350_cast_fp16 = add(x = q_49_cast_fp16, y = const_164_to_fp16)[name = string("op_1350_cast_fp16")];
tensor<fp16, [8, 64]> const_165_to_fp16 = const()[name = string("const_165_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70048320)))];
tensor<fp16, [1, 384, 8, 64]> var_1353_cast_fp16 = add(x = q_49_cast_fp16, y = const_165_to_fp16)[name = string("op_1353_cast_fp16")];
tensor<int32, [4]> q_with_bias_v_17_perm_0 = const()[name = string("q_with_bias_v_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
bool matrix_ac_17_transpose_x_0 = const()[name = string("matrix_ac_17_transpose_x_0"), val = bool(false)];
bool matrix_ac_17_transpose_y_0 = const()[name = string("matrix_ac_17_transpose_y_0"), val = bool(false)];
tensor<int32, [4]> transpose_46_perm_0 = const()[name = string("transpose_46_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
tensor<int32, [4]> transpose_47_perm_0 = const()[name = string("transpose_47_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
tensor<fp16, [1, 8, 64, 384]> transpose_47 = transpose(perm = transpose_47_perm_0, x = k_33_cast_fp16)[name = string("transpose_57")];
tensor<fp16, [1, 8, 384, 64]> transpose_46 = transpose(perm = transpose_46_perm_0, x = var_1350_cast_fp16)[name = string("transpose_58")];
tensor<fp16, [1, 8, 384, 384]> matrix_ac_17_cast_fp16 = matmul(transpose_x = matrix_ac_17_transpose_x_0, transpose_y = matrix_ac_17_transpose_y_0, x = transpose_46, y = transpose_47)[name = string("matrix_ac_17_cast_fp16")];
bool matrix_bd_33_transpose_x_0 = const()[name = string("matrix_bd_33_transpose_x_0"), val = bool(false)];
bool matrix_bd_33_transpose_y_0 = const()[name = string("matrix_bd_33_transpose_y_0"), val = bool(false)];
tensor<fp16, [1, 8, 64, 767]> var_1357_to_fp16 = const()[name = string("op_1357_to_fp16"), val = tensor<fp16, [1, 8, 64, 767]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70049408)))];
tensor<fp16, [1, 8, 384, 64]> q_with_bias_v_17_cast_fp16 = transpose(perm = q_with_bias_v_17_perm_0, x = var_1353_cast_fp16)[name = string("transpose_59")];
tensor<fp16, [1, 8, 384, 767]> matrix_bd_33_cast_fp16 = matmul(transpose_x = matrix_bd_33_transpose_x_0, transpose_y = matrix_bd_33_transpose_y_0, x = q_with_bias_v_17_cast_fp16, y = var_1357_to_fp16)[name = string("matrix_bd_33_cast_fp16")];
bool x_padded_33_interleave_0 = const()[name = string("x_padded_33_interleave_0"), val = bool(false)];
tensor<fp16, [1, 8, 384, 768]> x_padded_33_cast_fp16 = concat(axis = var_77, interleave = x_padded_33_interleave_0, values = (zero_pad_13_to_fp16, matrix_bd_33_cast_fp16))[name = string("x_padded_33_cast_fp16")];
tensor<int32, [4]> var_1374 = const()[name = string("op_1374"), val = tensor<int32, [4]>([1, 8, 768, 384])];
tensor<fp16, [1, 8, 768, 384]> x_padded_35_cast_fp16 = reshape(shape = var_1374, x = x_padded_33_cast_fp16)[name = string("x_padded_35_cast_fp16")];
tensor<int32, [4]> var_1378_begin_0 = const()[name = string("op_1378_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
tensor<int32, [4]> var_1378_end_0 = const()[name = string("op_1378_end_0"), val = tensor<int32, [4]>([1, 8, 768, 384])];
tensor<bool, [4]> var_1378_end_mask_0 = const()[name = string("op_1378_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
tensor<fp16, [1, 8, 767, 384]> var_1378_cast_fp16 = slice_by_index(begin = var_1378_begin_0, end = var_1378_end_0, end_mask = var_1378_end_mask_0, x = x_padded_35_cast_fp16)[name = string("op_1378_cast_fp16")];
tensor<int32, [4]> var_1383 = const()[name = string("op_1383"), val = tensor<int32, [4]>([1, 8, 384, 767])];
tensor<fp16, [1, 8, 384, 767]> shifted_17_cast_fp16 = reshape(shape = var_1383, x = var_1378_cast_fp16)[name = string("shifted_17_cast_fp16")];
tensor<int32, [4]> matrix_bd_35_begin_0 = const()[name = string("matrix_bd_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [4]> matrix_bd_35_end_0 = const()[name = string("matrix_bd_35_end_0"), val = tensor<int32, [4]>([1, 8, 384, 384])];
tensor<bool, [4]> matrix_bd_35_end_mask_0 = const()[name = string("matrix_bd_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
tensor<fp16, [1, 8, 384, 384]> matrix_bd_35_cast_fp16 = slice_by_index(begin = matrix_bd_35_begin_0, end = matrix_bd_35_end_0, end_mask = matrix_bd_35_end_mask_0, x = shifted_17_cast_fp16)[name = string("matrix_bd_35_cast_fp16")];
tensor<fp16, [1, 8, 384, 384]> var_1395_cast_fp16 = add(x = matrix_ac_17_cast_fp16, y = matrix_bd_35_cast_fp16)[name = string("op_1395_cast_fp16")];
fp16 _inversed_scores_33_y_0_to_fp16 = const()[name = string("_inversed_scores_33_y_0_to_fp16"), val = fp16(0x1p-3)];
tensor<fp16, [1, 8, 384, 384]> _inversed_scores_33_cast_fp16 = mul(x = var_1395_cast_fp16, y = _inversed_scores_33_y_0_to_fp16)[name = string("_inversed_scores_33_cast_fp16")];
tensor<fp16, [1, 8, 384, 384]> scores_35_cast_fp16 = select(a = var_59_to_fp16, b = _inversed_scores_33_cast_fp16, cond = mask_29_cast_fp16)[name = string("scores_35_cast_fp16")];
tensor<fp16, [1, 8, 384, 384]> var_1407_cast_fp16 = softmax(axis = var_77, x = scores_35_cast_fp16)[name = string("op_1407_cast_fp16")];
tensor<fp16, [1, 8, 384, 384]> input_191_cast_fp16 = select(a = var_68_to_fp16, b = var_1407_cast_fp16, cond = mask_29_cast_fp16)[name = string("input_191_cast_fp16")];
bool x_25_transpose_x_0 = const()[name = string("x_25_transpose_x_0"), val = bool(false)];
bool x_25_transpose_y_0 = const()[name = string("x_25_transpose_y_0"), val = bool(false)];
tensor<fp16, [1, 8, 384, 64]> v_35_cast_fp16 = transpose(perm = v_35_perm_0, x = v_33_cast_fp16)[name = string("transpose_60")];
tensor<fp16, [1, 8, 384, 64]> x_25_cast_fp16 = matmul(transpose_x = x_25_transpose_x_0, transpose_y = x_25_transpose_y_0, x = input_191_cast_fp16, y = v_35_cast_fp16)[name = string("x_25_cast_fp16")];
tensor<int32, [4]> var_1411_perm_0 = const()[name = string("op_1411_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<int32, [3]> var_1413 = const()[name = string("op_1413"), val = tensor<int32, [3]>([1, -1, 512])];
tensor<fp16, [1, 384, 8, 64]> var_1411_cast_fp16 = transpose(perm = var_1411_perm_0, x = x_25_cast_fp16)[name = string("transpose_56")];
tensor<fp16, [1, 384, 512]> input_193_cast_fp16 = reshape(shape = var_1413, x = var_1411_cast_fp16)[name = string("input_193_cast_fp16")];
tensor<fp16, [512, 512]> encoder_up_encoders_2_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_up_encoders_2_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70834880)))];
tensor<fp16, [512]> encoder_up_encoders_2_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_up_encoders_2_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71359232)))];
tensor<fp16, [1, 384, 512]> linear_62_cast_fp16 = linear(bias = encoder_up_encoders_2_self_attn_linear_out_bias_to_fp16, weight = encoder_up_encoders_2_self_attn_linear_out_weight_to_fp16, x = input_193_cast_fp16)[name = string("linear_62_cast_fp16")];
tensor<fp16, [1, 384, 512]> input_197_cast_fp16 = add(x = input_189_cast_fp16, y = linear_62_cast_fp16)[name = string("input_197_cast_fp16")];
tensor<int32, [1]> input_199_axes_0 = const()[name = string("input_199_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [512]> encoder_up_encoders_2_norm_ff_weight_to_fp16 = const()[name = string("encoder_up_encoders_2_norm_ff_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71360320)))];
tensor<fp16, [512]> encoder_up_encoders_2_norm_ff_bias_to_fp16 = const()[name = string("encoder_up_encoders_2_norm_ff_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71361408)))];
tensor<fp16, [1, 384, 512]> input_199_cast_fp16 = layer_norm(axes = input_199_axes_0, beta = encoder_up_encoders_2_norm_ff_bias_to_fp16, epsilon = var_58_to_fp16, gamma = encoder_up_encoders_2_norm_ff_weight_to_fp16, x = input_197_cast_fp16)[name = string("input_199_cast_fp16")];
tensor<fp16, [2048, 512]> encoder_up_encoders_2_feed_forward_w_1_weight_to_fp16 = const()[name = string("encoder_up_encoders_2_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71362496)))];
tensor<fp16, [2048]> encoder_up_encoders_2_feed_forward_w_1_bias_to_fp16 = const()[name = string("encoder_up_encoders_2_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73459712)))];
tensor<fp16, [1, 384, 2048]> linear_63_cast_fp16 = linear(bias = encoder_up_encoders_2_feed_forward_w_1_bias_to_fp16, weight = encoder_up_encoders_2_feed_forward_w_1_weight_to_fp16, x = input_199_cast_fp16)[name = string("linear_63_cast_fp16")];
tensor<fp16, [1, 384, 2048]> input_203_cast_fp16 = silu(x = linear_63_cast_fp16)[name = string("input_203_cast_fp16")];
tensor<fp16, [512, 2048]> encoder_up_encoders_2_feed_forward_w_2_weight_to_fp16 = const()[name = string("encoder_up_encoders_2_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73463872)))];
tensor<fp16, [512]> encoder_up_encoders_2_feed_forward_w_2_bias_to_fp16 = const()[name = string("encoder_up_encoders_2_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75561088)))];
tensor<fp16, [1, 384, 512]> linear_64_cast_fp16 = linear(bias = encoder_up_encoders_2_feed_forward_w_2_bias_to_fp16, weight = encoder_up_encoders_2_feed_forward_w_2_weight_to_fp16, x = input_203_cast_fp16)[name = string("linear_64_cast_fp16")];
tensor<fp16, [1, 384, 512]> input_209_cast_fp16 = add(x = input_197_cast_fp16, y = linear_64_cast_fp16)[name = string("input_209_cast_fp16")];
tensor<int32, [1]> query_axes_0 = const()[name = string("query_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [512]> encoder_up_encoders_3_norm_mha_weight_to_fp16 = const()[name = string("encoder_up_encoders_3_norm_mha_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75562176)))];
tensor<fp16, [512]> encoder_up_encoders_3_norm_mha_bias_to_fp16 = const()[name = string("encoder_up_encoders_3_norm_mha_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75563264)))];
tensor<fp16, [1, 384, 512]> query_cast_fp16 = layer_norm(axes = query_axes_0, beta = encoder_up_encoders_3_norm_mha_bias_to_fp16, epsilon = var_58_to_fp16, gamma = encoder_up_encoders_3_norm_mha_weight_to_fp16, x = input_209_cast_fp16)[name = string("query_cast_fp16")];
tensor<fp16, [512, 512]> encoder_up_encoders_3_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_up_encoders_3_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75564352)))];
tensor<fp16, [512]> encoder_up_encoders_3_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_up_encoders_3_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76088704)))];
tensor<fp16, [1, 384, 512]> linear_65_cast_fp16 = linear(bias = encoder_up_encoders_3_self_attn_linear_q_bias_to_fp16, weight = encoder_up_encoders_3_self_attn_linear_q_weight_to_fp16, x = query_cast_fp16)[name = string("linear_65_cast_fp16")];
tensor<int32, [4]> var_1457 = const()[name = string("op_1457"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 384, 8, 64]> q_55_cast_fp16 = reshape(shape = var_1457, x = linear_65_cast_fp16)[name = string("q_55_cast_fp16")];
tensor<fp16, [512, 512]> encoder_up_encoders_3_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_up_encoders_3_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76089792)))];
tensor<fp16, [512]> encoder_up_encoders_3_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_up_encoders_3_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76614144)))];
tensor<fp16, [1, 384, 512]> linear_66_cast_fp16 = linear(bias = encoder_up_encoders_3_self_attn_linear_k_bias_to_fp16, weight = encoder_up_encoders_3_self_attn_linear_k_weight_to_fp16, x = query_cast_fp16)[name = string("linear_66_cast_fp16")];
tensor<int32, [4]> var_1462 = const()[name = string("op_1462"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 384, 8, 64]> k_37_cast_fp16 = reshape(shape = var_1462, x = linear_66_cast_fp16)[name = string("k_37_cast_fp16")];
tensor<fp16, [512, 512]> encoder_up_encoders_3_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_up_encoders_3_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76615232)))];
tensor<fp16, [512]> encoder_up_encoders_3_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_up_encoders_3_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77139584)))];
tensor<fp16, [1, 384, 512]> linear_67_cast_fp16 = linear(bias = encoder_up_encoders_3_self_attn_linear_v_bias_to_fp16, weight = encoder_up_encoders_3_self_attn_linear_v_weight_to_fp16, x = query_cast_fp16)[name = string("linear_67_cast_fp16")];
tensor<int32, [4]> var_1467 = const()[name = string("op_1467"), val = tensor<int32, [4]>([1, -1, 8, 64])];
tensor<fp16, [1, 384, 8, 64]> v_37_cast_fp16 = reshape(shape = var_1467, x = linear_67_cast_fp16)[name = string("v_37_cast_fp16")];
tensor<int32, [4]> v_perm_0 = const()[name = string("v_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<fp16, [8, 64]> const_182_to_fp16 = const()[name = string("const_182_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77140672)))];
tensor<fp16, [1, 384, 8, 64]> var_1480_cast_fp16 = add(x = q_55_cast_fp16, y = const_182_to_fp16)[name = string("op_1480_cast_fp16")];
tensor<fp16, [8, 64]> const_183_to_fp16 = const()[name = string("const_183_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77141760)))];
tensor<fp16, [1, 384, 8, 64]> var_1483_cast_fp16 = add(x = q_55_cast_fp16, y = const_183_to_fp16)[name = string("op_1483_cast_fp16")];
tensor<int32, [4]> q_with_bias_v_perm_0 = const()[name = string("q_with_bias_v_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
bool matrix_ac_transpose_x_0 = const()[name = string("matrix_ac_transpose_x_0"), val = bool(false)];
bool matrix_ac_transpose_y_0 = const()[name = string("matrix_ac_transpose_y_0"), val = bool(false)];
tensor<int32, [4]> transpose_48_perm_0 = const()[name = string("transpose_48_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
tensor<int32, [4]> transpose_49_perm_0 = const()[name = string("transpose_49_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
tensor<fp16, [1, 8, 64, 384]> transpose_49 = transpose(perm = transpose_49_perm_0, x = k_37_cast_fp16)[name = string("transpose_52")];
tensor<fp16, [1, 8, 384, 64]> transpose_48 = transpose(perm = transpose_48_perm_0, x = var_1480_cast_fp16)[name = string("transpose_53")];
tensor<fp16, [1, 8, 384, 384]> matrix_ac_cast_fp16 = matmul(transpose_x = matrix_ac_transpose_x_0, transpose_y = matrix_ac_transpose_y_0, x = transpose_48, y = transpose_49)[name = string("matrix_ac_cast_fp16")];
bool matrix_bd_37_transpose_x_0 = const()[name = string("matrix_bd_37_transpose_x_0"), val = bool(false)];
bool matrix_bd_37_transpose_y_0 = const()[name = string("matrix_bd_37_transpose_y_0"), val = bool(false)];
tensor<fp16, [1, 8, 64, 767]> var_1487_to_fp16 = const()[name = string("op_1487_to_fp16"), val = tensor<fp16, [1, 8, 64, 767]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77142848)))];
tensor<fp16, [1, 8, 384, 64]> q_with_bias_v_cast_fp16 = transpose(perm = q_with_bias_v_perm_0, x = var_1483_cast_fp16)[name = string("transpose_54")];
tensor<fp16, [1, 8, 384, 767]> matrix_bd_37_cast_fp16 = matmul(transpose_x = matrix_bd_37_transpose_x_0, transpose_y = matrix_bd_37_transpose_y_0, x = q_with_bias_v_cast_fp16, y = var_1487_to_fp16)[name = string("matrix_bd_37_cast_fp16")];
bool x_padded_37_interleave_0 = const()[name = string("x_padded_37_interleave_0"), val = bool(false)];
tensor<fp16, [1, 8, 384, 768]> x_padded_37_cast_fp16 = concat(axis = var_77, interleave = x_padded_37_interleave_0, values = (zero_pad_13_to_fp16, matrix_bd_37_cast_fp16))[name = string("x_padded_37_cast_fp16")];
tensor<int32, [4]> var_1504 = const()[name = string("op_1504"), val = tensor<int32, [4]>([1, 8, 768, 384])];
tensor<fp16, [1, 8, 768, 384]> x_padded_cast_fp16 = reshape(shape = var_1504, x = x_padded_37_cast_fp16)[name = string("x_padded_cast_fp16")];
tensor<int32, [4]> var_1508_begin_0 = const()[name = string("op_1508_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
tensor<int32, [4]> var_1508_end_0 = const()[name = string("op_1508_end_0"), val = tensor<int32, [4]>([1, 8, 768, 384])];
tensor<bool, [4]> var_1508_end_mask_0 = const()[name = string("op_1508_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
tensor<fp16, [1, 8, 767, 384]> var_1508_cast_fp16 = slice_by_index(begin = var_1508_begin_0, end = var_1508_end_0, end_mask = var_1508_end_mask_0, x = x_padded_cast_fp16)[name = string("op_1508_cast_fp16")];
tensor<int32, [4]> var_1513 = const()[name = string("op_1513"), val = tensor<int32, [4]>([1, 8, 384, 767])];
tensor<fp16, [1, 8, 384, 767]> shifted_cast_fp16 = reshape(shape = var_1513, x = var_1508_cast_fp16)[name = string("shifted_cast_fp16")];
tensor<int32, [4]> matrix_bd_begin_0 = const()[name = string("matrix_bd_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
tensor<int32, [4]> matrix_bd_end_0 = const()[name = string("matrix_bd_end_0"), val = tensor<int32, [4]>([1, 8, 384, 384])];
tensor<bool, [4]> matrix_bd_end_mask_0 = const()[name = string("matrix_bd_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
tensor<fp16, [1, 8, 384, 384]> matrix_bd_cast_fp16 = slice_by_index(begin = matrix_bd_begin_0, end = matrix_bd_end_0, end_mask = matrix_bd_end_mask_0, x = shifted_cast_fp16)[name = string("matrix_bd_cast_fp16")];
tensor<fp16, [1, 8, 384, 384]> var_1525_cast_fp16 = add(x = matrix_ac_cast_fp16, y = matrix_bd_cast_fp16)[name = string("op_1525_cast_fp16")];
fp16 _inversed_scores_37_y_0_to_fp16 = const()[name = string("_inversed_scores_37_y_0_to_fp16"), val = fp16(0x1p-3)];
tensor<fp16, [1, 8, 384, 384]> _inversed_scores_37_cast_fp16 = mul(x = var_1525_cast_fp16, y = _inversed_scores_37_y_0_to_fp16)[name = string("_inversed_scores_37_cast_fp16")];
tensor<fp16, [1, 8, 384, 384]> scores_cast_fp16 = select(a = var_59_to_fp16, b = _inversed_scores_37_cast_fp16, cond = mask_29_cast_fp16)[name = string("scores_cast_fp16")];
tensor<fp16, [1, 8, 384, 384]> var_1537_cast_fp16 = softmax(axis = var_77, x = scores_cast_fp16)[name = string("op_1537_cast_fp16")];
tensor<fp16, [1, 8, 384, 384]> input_211_cast_fp16 = select(a = var_68_to_fp16, b = var_1537_cast_fp16, cond = mask_29_cast_fp16)[name = string("input_211_cast_fp16")];
bool x_transpose_x_0 = const()[name = string("x_transpose_x_0"), val = bool(false)];
bool x_transpose_y_0 = const()[name = string("x_transpose_y_0"), val = bool(false)];
tensor<fp16, [1, 8, 384, 64]> v_cast_fp16 = transpose(perm = v_perm_0, x = v_37_cast_fp16)[name = string("transpose_55")];
tensor<fp16, [1, 8, 384, 64]> x_cast_fp16 = matmul(transpose_x = x_transpose_x_0, transpose_y = x_transpose_y_0, x = input_211_cast_fp16, y = v_cast_fp16)[name = string("x_cast_fp16")];
tensor<int32, [4]> var_1541_perm_0 = const()[name = string("op_1541_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
tensor<int32, [3]> var_1543 = const()[name = string("op_1543"), val = tensor<int32, [3]>([1, -1, 512])];
tensor<fp16, [1, 384, 8, 64]> var_1541_cast_fp16 = transpose(perm = var_1541_perm_0, x = x_cast_fp16)[name = string("transpose_51")];
tensor<fp16, [1, 384, 512]> input_213_cast_fp16 = reshape(shape = var_1543, x = var_1541_cast_fp16)[name = string("input_213_cast_fp16")];
tensor<fp16, [512, 512]> encoder_up_encoders_3_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_up_encoders_3_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77928320)))];
tensor<fp16, [512]> encoder_up_encoders_3_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_up_encoders_3_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78452672)))];
tensor<fp16, [1, 384, 512]> linear_69_cast_fp16 = linear(bias = encoder_up_encoders_3_self_attn_linear_out_bias_to_fp16, weight = encoder_up_encoders_3_self_attn_linear_out_weight_to_fp16, x = input_213_cast_fp16)[name = string("linear_69_cast_fp16")];
tensor<fp16, [1, 384, 512]> input_217_cast_fp16 = add(x = input_209_cast_fp16, y = linear_69_cast_fp16)[name = string("input_217_cast_fp16")];
tensor<int32, [1]> input_219_axes_0 = const()[name = string("input_219_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [512]> encoder_up_encoders_3_norm_ff_weight_to_fp16 = const()[name = string("encoder_up_encoders_3_norm_ff_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78453760)))];
tensor<fp16, [512]> encoder_up_encoders_3_norm_ff_bias_to_fp16 = const()[name = string("encoder_up_encoders_3_norm_ff_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78454848)))];
tensor<fp16, [1, 384, 512]> input_219_cast_fp16 = layer_norm(axes = input_219_axes_0, beta = encoder_up_encoders_3_norm_ff_bias_to_fp16, epsilon = var_58_to_fp16, gamma = encoder_up_encoders_3_norm_ff_weight_to_fp16, x = input_217_cast_fp16)[name = string("input_219_cast_fp16")];
tensor<fp16, [2048, 512]> encoder_up_encoders_3_feed_forward_w_1_weight_to_fp16 = const()[name = string("encoder_up_encoders_3_feed_forward_w_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78455936)))];
tensor<fp16, [2048]> encoder_up_encoders_3_feed_forward_w_1_bias_to_fp16 = const()[name = string("encoder_up_encoders_3_feed_forward_w_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80553152)))];
tensor<fp16, [1, 384, 2048]> linear_70_cast_fp16 = linear(bias = encoder_up_encoders_3_feed_forward_w_1_bias_to_fp16, weight = encoder_up_encoders_3_feed_forward_w_1_weight_to_fp16, x = input_219_cast_fp16)[name = string("linear_70_cast_fp16")];
tensor<fp16, [1, 384, 2048]> input_223_cast_fp16 = silu(x = linear_70_cast_fp16)[name = string("input_223_cast_fp16")];
tensor<fp16, [512, 2048]> encoder_up_encoders_3_feed_forward_w_2_weight_to_fp16 = const()[name = string("encoder_up_encoders_3_feed_forward_w_2_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80557312)))];
tensor<fp16, [512]> encoder_up_encoders_3_feed_forward_w_2_bias_to_fp16 = const()[name = string("encoder_up_encoders_3_feed_forward_w_2_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82654528)))];
tensor<fp16, [1, 384, 512]> linear_71_cast_fp16 = linear(bias = encoder_up_encoders_3_feed_forward_w_2_bias_to_fp16, weight = encoder_up_encoders_3_feed_forward_w_2_weight_to_fp16, x = input_223_cast_fp16)[name = string("linear_71_cast_fp16")];
tensor<fp16, [1, 384, 512]> input_229_cast_fp16 = add(x = input_217_cast_fp16, y = linear_71_cast_fp16)[name = string("input_229_cast_fp16")];
tensor<int32, [1]> input_axes_0 = const()[name = string("input_axes_0"), val = tensor<int32, [1]>([-1])];
tensor<fp16, [512]> encoder_after_norm_weight_to_fp16 = const()[name = string("encoder_after_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82655616)))];
tensor<fp16, [512]> encoder_after_norm_bias_to_fp16 = const()[name = string("encoder_after_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82656704)))];
tensor<fp16, [1, 384, 512]> input_cast_fp16 = layer_norm(axes = input_axes_0, beta = encoder_after_norm_bias_to_fp16, epsilon = var_72_to_fp16, gamma = encoder_after_norm_weight_to_fp16, x = input_229_cast_fp16)[name = string("input_cast_fp16")];
tensor<fp16, [80, 512]> encoder_proj_weight_to_fp16 = const()[name = string("encoder_proj_weight_to_fp16"), val = tensor<fp16, [80, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82657792)))];
tensor<fp16, [80]> encoder_proj_bias_to_fp16 = const()[name = string("encoder_proj_bias_to_fp16"), val = tensor<fp16, [80]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82739776)))];
tensor<fp16, [1, 384, 80]> linear_72_cast_fp16 = linear(bias = encoder_proj_bias_to_fp16, weight = encoder_proj_weight_to_fp16, x = input_cast_fp16)[name = string("linear_72_cast_fp16")];
tensor<int32, [3]> var_1580_perm_0 = const()[name = string("op_1580_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
string cast_101_dtype_0 = const()[name = string("cast_101_dtype_0"), val = string("fp16")];
tensor<fp16, [1, 1, 384]> mask = cast(dtype = cast_101_dtype_0, x = masks)[name = string("cast_102")];
tensor<fp16, [1, 80, 384]> mu = transpose(perm = var_1580_perm_0, x = linear_72_cast_fp16)[name = string("transpose_50")];
} -> (mu, mask);
} |